Greg Isenberg Summaries & Notes

MetaMuse Opportunity

• MetaMuse (formerly Facebook Muse) offers an arbitrage opportunity to make money online by leveraging AI to mine public data from Facebook groups, Marketplace, and Instagram.
• The speaker believes this is a significant opportunity, potentially creating thousands of millionaires by connecting buyers and sellers more efficiently than traditional agencies.

How MetaMuse Works

• MetaMuse can identify needs and sources of supply within public online platforms.
• It cleans and organizes data, creating a valuable asset of contacts with specific needs and pricing.
• For enhanced functionality, MetaMuse can be integrated with services like Bland AI for a dedicated phone number, enabling calls and texts.

Monetization Strategies

• Service Arbitrage: Act as an intermediary, similar to a night nurse agency, by connecting those who need services with providers found via MetaMuse.
• Data Asset Creation: Build and own a clean, structured database of individuals and their needs, which could eventually be sold.
• Owning Platforms: Progress to owning the platforms themselves, such as creating and managing niche Facebook groups or Instagram pages.

Levels of Opportunity

• Level 1: Mining - Extracting valuable information from existing platforms.
• Level 2: Data Ownership - Cleaning and organizing mined data into a sellable asset.
• Level 3: Platform Ownership - Establishing and controlling your own online spaces (groups, pages).
• Level 4: Technology Building - Developing new technology or automated systems based on insights gained.

Undervalued Assets

• Facebook Groups: High user engagement makes them prime targets for MetaMuse.
• Facebook Business Pages & Instagram Niche Pages: Can be leveraged by MetaMuse for data collection.

Practical Application Example

• Using MetaMuse to find night nurses in a specific city by searching public Facebook groups and Instagram.
• The AI can then draft messages, make calls (with integration), and help vet potential candidates.

Vibe Marketing with Argon

• Argon excels at creating highly interactive and unique websites that stand out, moving beyond basic designs.
• To achieve impressive results, Argon requires good creative references.
• It can generate viral content for platforms like Twitter and TikTok, which can then be monetized by building a business around the generated audience.

Personalized Physical Products (VI Manufacturing)

• Argon can assist in designing personalized physical products, generating CAD files and embedding customizers into platforms like Shopify.
• The process involves customers customizing products (e.g., size, color) with a live 3D preview, followed by on-demand 3D printing via services like SLAN 3D or using a personal printer.
• This opens opportunities for niche physical products and even a service to build these customizers for other businesses.

Rebuilding Legacy SAS Products

• Argon can be used to build improved versions of older, less updated SAS products, particularly by niching down.
• A strategy involves offering a useful free tool to attract users, then developing paid tools for that user base.
• Argon can generate functional web applications, even for complex tasks like creating price proposals from site visit notes.

Maximizing Argon's Potential (Hacks)

• Plan: Agree on the build before starting to avoid constant file changes.
• Sub Agents: Use slash commands for parallel task execution (e.g., research, checking).
• Goal: Employ slash goal with reference images for better, albeit token-intensive, output.
• Browser: Utilize slash browser to test user journeys directly within the app.
• Learn: Turn corrections into reusable project rules or skills for improved future outputs.
• Mode: Default to "High" mode for generally better output quality.

Brag

• Brag is an AI skill that creates short, high-quality launch videos for projects or websites.
• It can generate videos with music, motion, and copy, offering various tones and formats (vertical, square, landscape).
• A potential use is creating videos for Product Hunt launches that lack them, offering a conversation starter for services.

Agent Reach

• Agent Reach helps AI agents gather information from social platforms like YouTube, Reddit, and X.
• It's useful for customer research, identifying pain points and tasks customers still perform manually despite using software.
• This can uncover opportunities for small tools or improvements to existing businesses.

Voice Studio

• Voice Studio is an open-source AI voice tool that can generate speech and dub videos into other languages.
• It can run locally on your hardware and is useful for creating multilingual product explanation videos or tutorials.
• There's potential for a business offering voice dubbing services for companies with international customers.

PPT Master

• PPT Master converts documents or notes into editable presentations, allowing for custom templates.
• It can be used to quickly generate proposals after sales calls or to create standardized materials for sales teams.
• The key is to write the presentation for someone who wasn't on the call, ensuring clarity and context.

Open Rig

• Open Rig allows the organization of coding agents into teams for collaborative tasks.
• Agents can pass work between each other, useful for processes like code changes and reviews.
• This is beneficial for client work, especially for post-launch additions and reducing manual instruction passing.

Free Tools

• Free Tools (by Corey Haynes) provides instructions and scripts for AI agents to create free tools as lead magnets.
• These tools can attract customers by addressing specific needs, like a packaging cost calculator for a supplier.
• The concept focuses on providing value upfront to attract and convert potential customers.

Grill Me

• Grill Me interviews users about an idea or decision, asking follow-up questions to uncover potential issues.
• It's valuable for pre-building analysis, especially for exciting ideas where crucial questions might be overlooked.
• The skill helps prepare by identifying assumptions and questions to take to potential customers for validation.

Impeccable

• Impeccable provides coding agents with specific design guidance to improve AI-generated interfaces.
• It offers commands like "critique" and "polish" to refine website aesthetics and user experience.
• This can be a service for businesses with poorly designed websites, offering specific, actionable design improvements.

AI Native Service Firms

• Leverage AI agents for most of the delivery, with humans reviewing exceptions.
• Example: Bookkeeping firms where agents prepare books and humans review exceptions.
• High margin, requiring fewer clients.
• Creative agencies offering "logos as a service" or "ads as a service."

Offline Businesses

• Focus on experiences that involve human interaction (e.g., communal saunas).
• Leverage AI for administrative tasks, but the core offering is the in-person experience.
• Start with a Minimum Viable Product (MVP) or "food truck" version to minimize initial costs.

Distribution

• Build and own organic audiences as the cost of reaching people increases.
• Utilize the Audience, Community, Product (ACP) funnel: build an audience, then a community, then a product.
• Meme pages, YouTube channels, and Instagram audiences can be valuable.

Proprietary Data Sets

• Data is highly valuable, especially data not readily available to large AI models.
• Opportunity for solo founders to gather and sell niche, useful information.
• A "gold rush" for acquiring unique data.

Domain-Specific Harnesses

• Harnesses are systems that provide AI models with tools, workflow rules, memory, and checks, tailored to specific industries.
• They are the "body and accessories" around the AI engine.
• Evolved from simple GPT wrappers, adding industry-specific functionalities.
• Examples include legal tech tools like Harvey.

Robotics and Physical AI

• Businesses utilizing robotics for physical tasks like window cleaning as a service.
• Opportunity not just in building, but also in renting or selling this equipment.

Physical Products with a Fan Following

• Create niche products for passionate communities (e.g., specialty coffee equipment).
• AI can assist in designing and prototyping hardware for underserved markets.
• Examples include minimalist phones and well-designed kitchenware.

Compute and Energy

• Businesses providing essential infrastructure for AI agents, such as chips, electricity, and cooling.
• High demand for compute power and energy.

Health, Longevity, and Care

• Focus on products and services for the aging population.
• Opportunities in caregiving services and longevity-focused ventures.

Marketplaces and Social Networks

• Platforms that connect buyers and sellers or facilitate social interaction.
• If successful, they create strong "moats" due to network effects.
• AI agents will likely seek to access these platforms.

Real Assets

• Physical assets like industrial workshops, warehouses, and real estate.
• Providing physical spaces for businesses and trades.

Vertical Agents

• AI agents assigned recurring tasks within a specific industry (e.g., purchasing for restaurants).
• Building specialized agents for niche industry needs.

Security

• Providing security solutions for the AI age, both software (e.g., password managers, network security) and physical (e.g., enhanced security services).
• Addressing increased risks of hacking and data breaches.

What are Forward Deployed Engineers (FDEs) and Why They Earn Millions

• FDEs deploy AI agents to drive significant value (e.g., $5-25M) for companies, earning a share of that value.
• This masterclass reveals concrete examples and workflows for becoming an FDE.

The Problem: Complex Business Processes

• Businesses are complex, with fragmented systems, multiple communication channels, and undocumented processes.
• "Applying AI" is not a simple fix; it requires deep process re-engineering.
• AI holding companies acquire traditional firms and embed FDEs to overhaul operations.
• Example: An accounting firm's process that looked simple was actually a multi-step, looping process with many exceptions.

The FDE Process: Mapping, Re-engineering, Building

• Step 1: Interviews (Human API): Understand undocumented knowledge held by employees.
• Step 2: Mining Systems of Record: Analyze data from CRM, ERP, etc., to infer processes.
• Step 3: Analyzing Existing Documentation: Review available documents (SharePoint, Drive, Slack, etc.).
• These three steps create a comprehensive picture of the company's operations.

Deploying Agents Strategically

• Classify process steps: Delete, Plain Code (deterministic), Agentic (judgment required), Human-in-the-Loop (risky tasks).
• Key Strategy: Deploy agents *within* existing systems of record (e.g., Salesforce, NetSuite) to minimize disruption and training. Avoid pitching AI-native replacements.
• For Private Equity firms, group portfolio companies by system of record to streamline deployments.

Valuing FDEs and Selling the Outcome

• FDEs provide value through cost savings, revenue uplift, and risk mitigation.
• Tailor pitches to the executive: CFOs focus on cost and efficiency (e.g., faster book closing), CROs on revenue, CHROs on talent acquisition/training.
• Quantify value with KPIs: reduced cycle time, increased straight-through rate, lower cost per transaction.
• The cost of agents is significantly lower than human labor, and they offer higher accuracy.

Key FDE Skills and Actionable Steps

• FDEs need expertise in understanding work, shipping production code (agents), and the AI layer (model selection, testing, governance).
• Personal Action Plan: Map your own complex systems, document 20 personal tasks, categorize steps, and then target SMBs for practice (offer free work initially).
• Focus on process re-engineering, not just slapping AI onto existing workflows.

OpenAI DevDay Key Announcements

• Dots: OpenAI's personal agent platform, aiming to be an app ecosystem for AI, with access to 1.2 billion weekly active users.
• Decisions API: Enables real-time, user-defined decision-making by focusing on specific questions with predefined answers.
• Agents API: Now supports computer use, allowing developers to build agents that interact with software to complete tasks.
• Login with ChatGPT: Enables users to sign in to third-party websites using their ChatGPT credentials and token allowance, opening thousands of business opportunities.
• GPT-4.5 Turbo: A major upgrade to GPT-4.5.
• DALL-E 3: Enhanced image generation capabilities.
• ChatGPT Team & Enterprise: New plans for businesses.
• Spaces: Collaborative cloud documents for AI agents.

Business Opportunities & Frameworks

• The new features create a significant opportunity for businesses to generate billions of dollars in transactions.
• Plugin Ecosystem: Developers can create plugins (mini-apps) that get selected by ChatGPT to fulfill user requests.
• "Login with ChatGPT" Strategy: Create free core apps, leverage existing ChatGPT subscriptions, and upsell advanced features. This lowers the barrier to entry for users.
• Niche AI Categories: Focus on building products for specific, underserved niches that weren't previously economically viable.
• Four-Step Framework for Building: Trigger, Decision, Action, Feedback. Businesses that own two of these stages are likely to be strongest.
• Real-World API: Connect vetted specialists in the physical world to handle tasks initiated through AI agents.
• Agent Discovery Analytics: Build analytics platforms to help plugins get discovered and optimized across various personal agent platforms.

The $5 Trillion Opportunity

• A $5 trillion opportunity exists over the next decade as millions of small businesses, owned by retiring individuals, go up for sale.
• Major investors like Thrive and General Catalyst are acquiring these businesses and using AI agents to significantly boost profitability.
• This presents a unique opportunity for solo founders and small teams, as large funds overlook smaller acquisition targets.

Why Now? The Convergence of Factors

• Owner Retirement: Approximately 1 million businesses, valued at $5 trillion, are expected to sell by 2035 due to owners retiring without successors.
• AI Capabilities: AI agents are now proficient enough to handle routine tasks in service businesses (e.g., data entry, chasing documents, status updates), often with greater accuracy than humans.
• Pricing Inefficiency: Service businesses typically operate on low profit margins (5-10%), making them undervalued and ripe for AI-driven margin expansion.

How Big Funds Are Doing It

• Thrive Holdings: Acquired nearly 50 local accounting firms, implementing AI to save accountants significant time (e.g., one firm saw a 31% time saving per accountant).
• General Catalyst: Invested heavily in companies like Long Lake (property management) and Crescendo (customer support), doubling margins and automating a high percentage of tasks.
• Acquisition Strategy: Funds often pay 60-70% cash, with founders rolling over equity, incentivizing them to ensure a smooth transition.
• Playbook: Build AI platforms first, acquire trusted firms, integrate AI agents discreetly to improve efficiency, then scale by acquiring and integrating more firms.

The Solo Founder Approach

• Holding Company Structure: A solo founder can operate a holding company that owns multiple small businesses, each with a General Manager (GM) and a shared AI agent platform.
• Scalable Platform: The core AI agents and back-office systems are built once and reused across all acquired businesses, making subsequent acquisitions easier.
• Access to Tools: Solo founders can leverage the same AI models (e.g., CodeX, Gemini) as large funds.
• Integration Edge: Being the direct integrator allows for better client understanding and trust compared to large, impersonal funds.
• Acquisition Niche: Focus on smaller deals (<$2M revenue) that are too small for large funds, reducing competition.

Operationalizing AI Rollups

• Folder Structure: Organize operations with folders for Thesis, Shared Agents/Rules, Business-Specific Rules, Client Files, People Files, and a Corrections Log.
• Agent Workflow: Intake Agent (collects/chases), Prepare Agent (first draft), Reviewer Agent (checks against rules). Human approval is always required before client delivery.
• Agent Job Descriptions: Clearly define agent capabilities and limitations to prevent errors and ensure human oversight.
• Weekly Schedule:
  • Monday: Review key metrics (profit margin, human time per job, agent fix frequency, client retention, staff happiness).
  • Tuesday: Calls with GMs to discuss operations and challenges.
  • Wednesday: Analyze the Corrections Log to create new rules and improve agent performance.
  • Thursday/Friday: Prospect for new business acquisitions.
• Acquiring the First Business: Start by offering AI-powered services to businesses in a chosen industry (e.g., bookkeeping), build trust, and eventually acquire one of those clients.

Addressing Objections

• Rollups Fail: Traditional rollups fail due to overpaying or poor integration. AI rollups mitigate this by buying slowly, integrating personally, and pricing based on current value.
• Margin Competition: While AI adoption will spread, there's a significant time lag before margins are fully competed away. Strong client relationships and proprietary systems offer sustained advantages.
• Regulated Industries: Human oversight is crucial, but AI drafts significantly reduce the time needed for review, maintaining profitability.
• Resistance to Change: Minimize visible changes initially, run agents in the background, and incentivize key staff (like GMs) to ensure smooth transitions.
• Layoffs: Roles will shift from execution to oversight. While some job displacement is possible, the focus is on augmenting human capabilities and increasing productivity.

Meta Muse AI Connectors Launch

• Meta has opened its personal AI agent, Muse, to developers, allowing them to submit "connectors."
• Connectors enable Muse to use third-party services to fulfill user requests.
• This is compared to Apple's 2008 App Store launch, potentially creating an "app store moment" for AI.

Understanding Muse Connectors

• Connectors act as APIs, allowing Muse to interact with external services.
• Services can become useful mid-request, even if not the initial trigger.
• A submission process and directory for approved connectors exist.

Muse Connector Business Opportunities

• Lead Generation for Business Suppliers: Connecting businesses (e.g., linen services) with new local opportunities (e.g., restaurants opening).
• Home Repair Dispatch: Matching users with broken appliances to specific repair services with confirmed availability.
• Local Paddle Match/Court Finder: Connecting players with games and courts, leveraging Meta's social focus.
• Family Dinner Planning & Grocery Integration: Creating meal plans and grocery lists, with potential integration with services like Instacart.

Marketing and Growth Strategies

• Leverage existing creators with relevant audiences to demonstrate connector utility.
• Implement product-driven viral sharing and product-led growth (e.g., shareable booking details).
• Distribute through existing connected marketplaces or Meta's featured directory (though the latter is less reliable).

Building and Submitting a Connector

• Start by defining a basic customer task (e.g., "show me available courts").
• Utilize coding agents (e.g., Cloud Code, Codex) with API documentation and plain language descriptions.
• Ensure services can handle awkward requests and edge cases for smoother approval.
• Test thoroughly and understand the service's capabilities before customer use.
• Meta requires approval for connectors, with submission forms asking for product details and usage examples.

Introduction to AI Native Service Firms

• The rise of one-person businesses powered by AI, specifically using Astra and creative models.
• These firms offer services ranging from $500 to $5,000+, found on platforms like Upwork.
• Key enablers: advanced creative models, Astra for orchestration, and unified APIs (like Hicksfield) for accessing multiple models.

AI Native vs. Traditional Agencies

• Traditional agencies have lower net margins (10-15%) due to human labor costs and recruitment challenges.
• AI native firms leverage AI for fulfillment, enabling higher margins (60%+) and access to new sales channels (e.g., Upwork, Fiverr).
• AI native firms create a continuous learning loop, improving services over time.

Building an AI Native Service Firm with Studio Operator

• Studio Operator acts as a cockpit for managing AI creative service work.
• It integrates client requests, AI-driven production, quality assurance, and delivery.
• The system allows for defining budgets, repair limits, and approved models to maintain profitability.
• Astra manages the workflow, extracting data from client briefs and orchestrating model usage.

Key AI Models and Their Applications

• Soul: Best for fashion, editorial, portraits, founder photoshoots (weak on packaging, logos, copy).
• Cadream/Flux: Ideal for product placement in lifestyle images, ensuring subject recognition.
• Ideogram/Recraft: Suitable for text and layout needs like posters, covers, campaign graphics.
• Quen: Effective for repairing specific details like labels, headlines, or minor visual errors.
• Z Image/Marketing Studio Image: Useful for cheap, rapid exploration of still image concepts.
• Pixver: For quick video exploration of hooks, openings, and camera ideas.
• C-Dance: Best for longer cinematic scenes, multi-reference sequences, and native audio.
• Clang: For hero shots, camera movement, physical realism, and final polish.
• Juan/Miniaax: For speaking roles, creative ads, and spokesperson videos (watch for lip-sync).
• Cling Motion Control: For applying specific movements to characters based on reference.
• Topaz/Finishing Tools: For upscaling, localization, voice, captions, and final file export.

Startup Ideas for AI Native Service Firms

• Always-On Ad Studio: Weekly creative packs for e-commerce brands, using AI to generate hooks, scenes, and ads.
• Franchise Localization Service: Creating 50+ local versions of national campaigns with specific addresses, offers, and languages.
• Industrial Catalog Sales Team: Transforming dense PDFs and specs into product demos, application videos, and sales graphics for manufacturers.

Filters for Identifying Strong AI Business Opportunities

• Existing Budget: Customers are already paying for the service.
• Obvious Unit: Clearly defined deliverables (e.g., one product video, one campaign).
• Repetition: The work is repeatable, allowing for workflow optimization and subscription models.
• Concrete Definition of Correct: Objective criteria exist to verify the output (e.g., product name, claims, formats).

The Future of AI Native Businesses

• AI native firms build a continuously improving learning loop powered by data and workflows.
• Astra combined with specialized creative models enables the creation of highly capable, semi-autonomous service firms.
• Success relies on focusing on a narrow package, securing initial customers, and translating lessons learned into improved software workflows.

The AI Era and Content Value

• AI makes generic content worthless; a "point of view" creates a moat in SaaS and content.
• Language is open-source; the value lies in personal association with ideas, not the words themselves.
• AI remixes existing "approved language," not true original thought.

Building Your "Digital Brain"

• Develop a library of your stories, opinions, and lessons to build trust and customer relationships.
• This library acts as your "personal language model" or "digital brain."
• Focus on creating additive value with each piece of content to your digital brain.

Commodity vs. Personality Content

• Commodity content: Widely known, repeated ideas (e.g., "eat healthy to lose weight").
• Personality content: Content driven by your unique perspective, experiences, and "I" statements (e.g., "I think this would be a great startup idea").
• AI struggles with personality content because it lacks your unique data set.

Leveraging AI and Content Strategy

• Use AI for repetition and remixing of your existing content, not for generating new ideas.
• Focus on building your "unmade data set" – your personal experiences and formative moments.
• Writing creates association and ownership, which AI can amplify by repeating your established voice and ideas.

Voice, Style, and Timelessness

• Voice is a combination of word choice, sentence structure, and rhythm, intentionally created.
• Most content is 80% commodity language with small subjective variations; don't let voice concerns paralyze you.
• Evaluate content on a spectrum of timely (short-term value, e.g., memes) vs. timeless (long-term value, e.g., essays).

What is Jev?

• Jev is a new type of AI, a classifier AI, created by Dooo Almeida, the founder of ChatGPT.
• It is fundamentally different from LLMs like ChatGPT in terms of quality, speed, and price.
• Jev acts as an AI decision-maker, providing probabilities for choices rather than definitive answers.
• It does not generate text; instead, it takes an input and a schema, returning probabilistic outputs.

How Jev Works

• You define an input (e.g., an email object) and an output schema (e.g., categories, priority, spam score, reply percentage).
• Jev classifies the input based on the schema and returns confidence scores (percentages) for each option.
• Example: For an email, Jev can determine its category (shopping, work, marketing) and spam score (e.g., 90% spam).
• It's designed for tasks requiring quick, decision-based classification, not conversational interaction.
• Outputs are type-safe, allowing instant integration into code.
• Jev queries are extremely fast, around 200 milliseconds per query.

Jev Use Cases and Startup Ideas

• Email Triage: Instantly categorize and prioritize emails, identifying leads and urgent issues.
• Lead Qualification: A graphic design agency uses Jev to classify inbound leads based on quality.
• Customer Support: Automatically route support inquiries to the correct product team.
• Service Aggregation: Match users with local service providers (e.g., power washing) based on input criteria.
• Instant Quoting: Provide immediate quotes for services instead of delaying responses.
• Content Clipping: Transcribe videos and use Jev to identify and extract key moments for short-form content.
• Browser Control: Demonstrated controlling a browser to find a flight in 7.1 seconds.

Accessibility and Cost

• Jev is currently invite-only but accessible via the Vercel Gateway.
• It is incredibly fast and cheap, with low barriers to entry.
• The cost for processing 1,700 emails was only 18 cents.
• Testing is highly encouraged due to its low cost and speed.

Instinct AI Overview

• Instinct AI is an invite-only personal agent app with a rumored $10 billion valuation.
• It aims to be a personal agent for everyone, including non-technical users.
• The app excels in its natural conversational style, use of reactions, and integration with Apple iMessage features.
• It can even play iMessage games like Cuppong.

Key Features and Use Cases

• Instinct AI remembers context from previous conversations to inform current tasks.
• It proactively assists users, like reminding them about order cut-off times.
• The app simplifies complex tasks such as filling out visa-on-arrival forms and booking appointments.
• It offers a dedicated email address for the agent to manage bookings and tasks without cluttering the user's inbox.
• Users can loop the agent into email threads for task completion.
• A "Vault" feature stores login details, credit cards, and addresses for seamless task execution.

Onboarding and User Experience

• Onboarding is exceptionally simple, requiring only a phone number for iMessage integration.
• The interface is intuitive and designed for mass-market adoption, unlike more complex agent platforms.
• The app uses native Apple iMessage effects like confetti and loud messages to feel more human.
• It demonstrates tasks like booking flights, ordering food, and cancelling subscriptions during onboarding.

Privacy and Security Concerns

• Concerns exist regarding Instinct AI indexing and retaining user emails without explicit permission.
• Terms and conditions reportedly contain clauses about retaining control of private data.
• Users are advised to consider these privacy implications if they are a concern.

Network Effects and Future Potential

• The invite-only approach, reminiscent of Clubhouse, creates exclusivity and hype.
• A "Trusted Person Network" feature allows instinct agents to communicate with each other, representing a new form of network effect for the agentic era.
• Instinct AI is positioned as ideal for personal assistant tasks, while business/coding tasks might be better suited for platforms like Hermes or OpenClaw.

What is a Software Factory?

• A software factory uses AI agents like an assembly line to ship high-quality software efficiently.
• It's a workflow, skill set, and domain knowledge packed into specific development processes.
• It is harness and model agnostic, focusing on the process, not specific tools.

Why Software Factories Matter

• Intelligence of AI models is increasing, and a software factory systemizes their use for efficiency.
• It allows for faster shipping of valuable software without sacrificing quality.

The Four Steps of a Software Factory Workflow

• Step 1: Isolate (New Feature Skill)
  • Creates a new Git work tree branched from origin/main for each feature.
  • Ensures agents work in parallel without conflicts, preventing overwriting issues.
  • Allows for simultaneous development of multiple features by different agents or individuals.
• Step 2: Build (Code Structure Skill)
  • Agents write code following a service layer architecture for better readability and maintainability.
  • This makes it easier for humans or other agents to understand and work with the code.
• Step 3: Prove (Evidence-Driven Testing / Before & After Skills)
  • Agents must prove their work is done correctly.
  • Evidence-driven testing records before/after states (videos or screenshots) to verify fixes or new features.
  • If proof is not visual (e.g., performance improvements), agents provide test results.
  • This step builds trust and allows for easier review, especially for non-technical individuals.
• Step 4: Ship (GP Loop / Code Review Agents)
  • Utilizes external code review agents (like Grapile) with skills like GP Loop.
  • Agents receive feedback from review tools and automatically loop back to the 'Build' step if issues are found.
  • This iterative process continues until a high confidence score (e.g., 5/5) is achieved.
  • Once quality is assured, the code is merged into the main branch.

Key Takeaways

• A software factory is a methodology, not a product or specific tool.
• It involves using markdown files to define workflows and agent behaviors.
• The process emphasizes isolation, structured building, rigorous proof, and automated review loops.
• Using code review agents is highly recommended for serious software development to ensure quality and user empathy.

Astra Model Overview

• GPT-6 Astra is OpenAI's most expensive model, priced comparably to Fable.
• It's efficient, requiring fewer steps to complete tasks.
• Subsidized pricing is available with a Codex subscription.

Productivity & Business Applications

• Performance Updates: Improve app/API speeds significantly (e.g., 800ms to 20-30ms).
• Security Audits: Identify vulnerabilities in existing applications.
• UI Design: Generate user interfaces, especially effective with a reference image.
• Bill Renegotiation: Automate negotiation with service providers to lower costs.
• Service to Software: Reverse engineer service workflows to create AI products for recurring revenue ($500-$5,000/month).
• Prototyping: Rapidly build clones to explore app ideas.
• Deal Hunting: Monitor marketplaces for underpriced items (e.g., cameras, furniture).
• Founder Dashboard: Consolidate docs, notes, Stripe data, calls, etc., for a weekly operational overview.
• Agent Opportunity Audit: Identify tasks suitable for AI agents before hiring new staff.
• Browser Operator: Automate complex web tasks, data collection, and form filling.
• QA Team: Automate app testing on real devices, identifying bugs and UX confusion.
• Competitor Spy: Generate monthly reports on competitors' new features and pricing.
• Viral Game Development: Create browser games with lead capture and viral sharing mechanics.

Hardware & Advanced Projects

• AI Home Speaker: Build a DIY AI home speaker using Raspberry Pi and a custom agent.
• 3D Prototyping: Use Blender with Astra to visualize hardware assembly and design.
• Hardware Sourcing: Obtain links for necessary components and identify Chinese suppliers for shells.
• Code Generation: Astra can write and even create pull requests for agent integration code.
• PS4 Linux Lab: Repurpose a PS4 for running Linux and experimenting with agents.
• Vibe Manufacturing: Ushering in an era of rapid hardware prototyping and niche product creation.

Value & Call to Action

• Astra enables ambitious projects that were previously intimidating or required specialized skills.
• The current subsidization offers immense value, potentially saving users thousands in API costs.
• Encouragement to "dream big, act big, and build big" with Astra.
• Suggested to cancel other AI subscriptions (like Claude) in favor of Astra's capabilities.

What is Local AI?

• Local AI means the AI model runs on hardware you control (laptop, phone, etc.), unlike cloud AI which runs elsewhere.
• Local AI is ideal for tasks involving private/sensitive data, offline usage, low latency, or repetitive internal workflows.

The Local AI Landscape

• Model: The "brain file" (e.g., Gemma, Llama, Mistral).
• Warehouse: Where models are found (e.g., Hugging Face).
• Software: Tools to run models (e.g., LM Studio, O Lama).
• Workflow: The product built around the model.

Key Concepts in Local AI

• Parameters: Model size; more parameters generally mean more capacity but require more memory/speed.
• Tokens: Text chunks the model reads/writes; focus is on speed/memory, not per-token cost.
• Context Window: How much information the model can process at once.
• Quantization: Compressing models to run on less powerful hardware (e.g., Q4, Q8 formats).
• GGUF: A common file format for local models.
• Light RTLM: Google's runtime for on-device AI apps.

Google's Gemma Models & AI Edge

• Gemma is Google's open model family, with Gemma 4 optimized for local/on-device use.
• Gemma 4 E4B is a practical starting point for local tests.
• Specialized Gemma models include Embedding Gemma (for search) and Function Gemma (for tool use).
• Google AI Edge is the broader on-device AI development ecosystem.

Running Models Locally

• LM Studio: User-friendly desktop app for searching and chatting with models.
• O Lama: More developer-oriented; allows running models locally with an API.
• Google AI Edge/Light RTLM: For building actual on-device AI products.

Hardware Considerations

• RAM is key: 8GB for small models, 16GB for E4B, 32GB+ for larger models.
• Phones are suitable for specific tasks like image understanding or audio summarization.

Building with Local AI: A Workflow Approach

• Start with existing workflows and private data folders (e.g., customer notes, support tickets).
• Use local AI to generate useful artifacts (memos, checklists, reports) rather than just chat answers.
• Compare local model output to cloud models to understand hybrid approaches.
• Prioritize local for private data, offline use, device-native tasks, and repetitive workflows.
• Use cloud for deep reasoning, large context, and when the strongest model significantly improves quality.

Startup Ideas with Local AI

• Idea 1: Local QA reviewer for home health agencies (reviews visit notes for compliance/billing issues).
• Idea 2: Offline field report co-pilot for restoration contractors (mobile app drafting reports on-site).
• Idea 3: Local pre-send reviewer for professional services (flags sensitive info or errors in client drafts).

No AI Slop Skill

• Addresses the problem of AI-generated writing having a distinct, often undesirable, "AI smell."
• Helps preserve human voice while removing AI patterns, making content more believable.
• Workflow: Write a draft, then use the skill to refine it, keeping human ideas primary.
• Installation: Use `npx skills add [GitHub link]` to install as a skill.

CRM by TryCompai

• Open-source CRM designed for AI agents, turning a static database into a dynamic workspace.
• AI agents can research, enrich data, and manage follow-ups automatically.
• Aims to maintain relationships and prevent opportunities from being lost, a common cause of business failure.
• Use cases: Sponsorship pipelines, agency leads, investor updates, waitlist follow-ups, small SaaS customer success.
• Setup requires Bun, Docker, and environment variable configuration for Oauth.

Video Editing with Code Agents (Video Use)

• Enables editing raw footage using AI agents (e.g., Claude Code, Codex).
• Capabilities include removing filler words, cutting dead space, adding subtitles, color grading, and rendering.
• Makes editing workflows explicit and repeatable for agents.
• Workflow: Start with small, repeatable formats (e.g., turning Looms into short videos).
• Installation: Agents can often install it via a prompt, or manual installation via Git clone is possible. Requires FFmpeg.

Skill Spectre (Nvidia)

• Scans AI agent skills for security vulnerabilities like prompt injection and data exfiltration.
• Crucial for securing modular AI workflows and preventing malicious code execution.
• Provides a safety check before integrating new capabilities into an AI agent setup.
• Installation: `uv tool install git+ [GitHub link]` and then `skillspector scan [path/repo]`. Supports `--no-lm` for static scans.

Phone Harness

• Allows AI agents to control real iPhones and Android phones.
• Enables automation of mobile app workflows, including testing, sign-ups, and checkouts.
• Uses iPhone mirroring for iOS and ADB for Android. No jailbreaking required.
• Potential for mobile QA services, automating apps without APIs, and repetitive creator/operator tasks.
• Setup involves cloning the repo, installing, and granting necessary permissions. Early stage with some limitations (e.g., lock screens, camera flows).

The Rise of the Marketing Engineer

• The "Marketing Engineer" (also called Forward Deployed Marketer or AI Growth Operator) is a new, highly valuable role combining marketing with AI agents.
• This role can command salaries of $250k-$1M due to companies' need for more leads, faster experiments, sharper positioning, and smarter marketing with smaller teams.
• Marketing shifts through eras: Don Draper (storytelling), Digital Marketer (measurable channels), Growth Hacker (product loops), and now Marketing Engineer (AI, agents, data, code, taste).

The Marketing Engineer's Role and Workflow

• Marketing Engineers build systems that learn by connecting customer data, analyzing results, shipping landing pages, and converting market signals into actionable insights.
• The core job: Turn market signal into pipeline using AI agents, data, code, taste.
• Key output: A "Growth Repo" (like Growth OS) acting as marketing memory, containing:
  • Customer Truth (sales calls, support tickets, churn notes)
  • Content Engine (founder voice, winning hooks, performance notes)
  • Outbound Engine (ICP, trigger events, approved language)
  • Creative Testing (ad angles, landing page tests)
  • Agents (defined AI worker jobs)
• Prompts become highly contextual, e.g., "Read customer truth and founder voice files, draft posts based on pain points mentioned this week."

Essential Tool Stack and Concepts

• Grockbot: An "operating system" close to the internet, useful for monitoring competitors, customer language, creators, and ads.
• Claude/Codeex: Used for building the repo, generating content, scripts, and internal tools.
• Hermes-style workflows: For scheduled operations with memory and approvals (e.g., weekly market briefs).
• Creative Models (e.g., Foul AI, Higsfield): For generating ad creatives, thumbnails, mockups, video concepts.
• Local AI: For handling sensitive or expensive data.
• Agents need clear "job specs": data sources, run times, filters, expected output, approval criteria, and key metrics (e.g., positive replies from qualified accounts, not just messages sent).
• Training agents is like training new hires: start small, correct mistakes, add corrections to memory, expand scope.

Concrete Example: Vertical SAS for HVAC Contractors

• Problem: Identifying specific pain points that drive demos (e.g., "lost replacement revenue" vs. "run business better").
• System 1: Customer Truth System: Generates a `what_the_market_is_telling_us.md` file from sales calls, support tickets, etc., highlighting changes in market language or user sticking points with specific quotes and links.
• System 2: Founder Content Engine: Extracts strong ideas from founder insights and customer stories, using performance data to create content (posts, videos, landing pages, emails).
• System 3: Outbound Signal Engine: Monitors timing signals (funding, hiring) to identify accounts for targeted outbound messages.
• System 4: Creative Testing Engine: Spins up and tests numerous ad angles and hooks based on positioning.
• System 5: AI Search Visibility: Optimizes content for AI search engines (ChatGPT, Gemini, Perplexity).
• System 6: Growth Cockpit: A weekly dashboard showing performance metrics, competitor moves, and actionable insights for next steps.

Making Money as a Marketing Engineer

• Become the In-House Expert: Increase value within a company by driving pipeline, conversions, and reducing wasted spend. Best marketing engineers can drive insane value, leading to high compensation.
• Consulting: Embed with companies for short periods (30-90 days) to build specific growth systems (e.g., Customer Truth System, Outbound Engine) and charge for outcomes.
• Productized Services: Specialize in a niche offering (e.g., Outbound Signal Engines for Vertical SAS) and repeat the process for efficiency and scalability.
• Software: Build tools/agents after identifying repeating pain points from service delivery. Start with services to validate the need.

30-Day Plan to Become a Marketing Engineer

• Week 1: Audit: Choose a company, study its website, offer, ICP, content, and available data (sales calls, tickets). Create a market map outlining customer pain, language, funnel leaks, and initial test ideas.
• Week 2: Growth Repo: Create the repo, set up folders, and build the "What is the Market Telling Us" file using chosen AI tools. Focus on transforming scattered signals into meta-insights with receipts.
• Week 3: Build First Machine: Select and build one core system (e.g., Content Engine, Outbound Engine, Landing Page Tester). Prioritize completing one working system over multiple unfinished ones.
• Week 4: Results: Measure changes in replies, meetings booked, conversion lift, or founder satisfaction. Document learnings and business results (e.g., "Built Customer Truth Repo, identified high-intent pain, created outbound engine shipping X messages, booking Y calls").

The Future and Key Skills

• Marketing Engineers blend marketer, product person, RevOps, data analyst, creator, and engineer skills.
• Their "moat" is judgment and taste in directing AI agents, not just the agents themselves.
• This role offers a significant edge over those stuck in older marketing paradigms. The window of opportunity is now.

What is WebMCP?

• WebMCP is a new experimental feature by Google and Microsoft that allows websites to expose cleaner actions for AI agents.
• It enables websites to tell AI agents how to search, book, or buy, making them "agent ready" and "agent actionable."
• WebMCP makes websites agent-readable and agent-actionable, a step beyond SEO and AEO (AI-enhanced optimization).

How WebMCP Works

• Instead of agents scanning the entire DOM (website code) or taking screenshots, WebMCP provides a concise list of tools agents can use.
• These tools can be conditional, appearing only when a user is logged in, leveraging the browser session for authentication.
• This is a more efficient and less fragile method compared to traditional web scraping or headless approaches.

Benefits and Opportunities

• WebMCP bridges the gap between new AI agent technology and users accustomed to visual interfaces, facilitating wider adoption of AI tools.
• It simplifies agent integration by using existing browser sessions and login credentials, avoiding the complexities of API keys and tokens.
• Opportunities lie in e-commerce, SaaS admin, regulated industries (read-only tools), and internal company tools.

Startup Ideas

• WebMCP Conversion Agency: Make traditional business websites (law firms, home services) agent-ready by adding V1 tools for tasks like quote requests or consultations, offering setup fees and monthly retainers.
• Agent Mystery Shopper: Test and report on how well AI agents can complete key user journeys on websites, identifying conversion risks and areas for improvement, with potential to build software from repeated fixes.

Getting Started

• WebMCP is experimental; enable it in Chrome flags and allow remote debugging in Chrome inspect.
• A demo site is available at crema.co/clientfly.dev, and the GitHub repo can be cloned for experimentation.
• Early adoption of experimental features like WebMCP offers a first-mover advantage.

Introduction to Grokbot

• Grokbot is a new AI agent team tool enabling non-technical users to build and run businesses end-to-end.
• It allows for the creation of AI agent teams to run businesses, with potential earnings of $2k-$25k/month.
• Grokbot is presented as a user-friendly, non-technical alternative to tools like Jack Dorsey's Buzz product.

Grokbot's Unique Approach and Best Practices

• Grokbot's constraint-based design (limited agent types, custom avatars) encourages mission-oriented projects.
• It offers a Slack-like interface, differentiating agents by color and shape, reducing switching costs and improving organization.
• Best Practice: Focus on one project at a time to avoid token debt and context bloat.
• Grokbot agents run on the cloud, potentially on virtual machines, consolidating operations.

Setting Up and Managing Grokbot Agent Teams

• Start by setting up a "Chief of Staff" agent to audit existing business documents (Notion, Slack, Gmail).
• The Chief of Staff should identify the top three agents needed to drive revenue.
• For the Arlington Bagel newsletter example, the initial agents were: Research, Chief of Staff (editing), and a Beehive expert.
• Key Principle: The human user must make critical decisions; don't let agents get stuck in endless analysis.
• Execution Framework:
• Week 1: Build the team and understand its capabilities.
• Week 2: Execute and avoid tinkering ("no tinkering").
• Week 3: Expand the team based on identified gaps (e.g., Gmail monitoring, Shopify shop management).
• Week 4: Implement automation through "routines."

Automation and Advanced Techniques

• Routines can be set up for daily or weekly tasks, generating end-of-day briefs.
• Example: A social media agent creates graphics and content calendars, reporting blockers and progress to the Chief of Staff.
• Adversarial reviews (QA loops) can be automated by using sub-agents to critique work, improving output quality.
• Grokbot can be trained to mimic specific review styles or expert feedback.

Business Ideas and Monetization with Grokbot

• Local Newsletters: Research local news, events, and restaurants. Monetize through sponsorships.
• Niche Newsletters: Applicable to any niche (tech, sports, board games).
• Directories: Build curated data websites for SEO and affiliate marketing.
• Shopify Stores: Agents can research products, sourcing, and use developer tools (like Shopify CLI).
• Internal Tool Building: Create sites for specific needs (e.g., meeting booking pages, landing pages for local businesses).
• Service Businesses: Build directories and then monetize through lead generation for sponsorships or e-commerce.

Technical Integration and Tips

• Grokbot supports plugins for integrating with platforms like Beehive and Notion.
• For platforms without direct integration (e.g., Shopify), agents can be trained to use command-line tools (CLI).
• Grokbot can access local files, but users should direct output to chat to avoid organizational issues.
• Using tools like Make.com with OpenAI can be more cost-effective for high-volume, repetitive tasks like writing short blurbs, freeing Grokbot tokens for core business building.
• When researching, provide specific tasks (e.g., "recommend top 3 ideas") rather than open-ended prompts.

Conclusion and Call to Action

• Grokbot offers a powerful yet simple way to build businesses with AI agents.
• The recommended approach is to focus on one project, iterate, and automate.
• Users are encouraged to try Grokbot for a month on a chosen project (newsletter, directory, Shopify) and share their progress.

What are AI Skills?

• Skills are Standard Operating Procedures (SOPs) for AI, packaging specific instructions and preferences into a markdown file.
• They enable AI agents to perform tasks consistently and accurately without repetitive re-explanation.
• A good skill can save significant time (e.g., 2 hours/week).

The Challenge of Sharing Skills

• Currently, AI usage is often "single-player"; skills created by individuals are isolated on their machines.
• Sharing skills via simple file transfer leads to version control issues and a lack of a central source of truth.
• Traditional cloud storage (Drive, Dropbox) and note-taking apps (Obsidian) require complex setup and don't integrate seamlessly with AI tools.

The GitHub Plugin Solution

• The most effective way to share and manage skills across a team is by using a GitHub repository and turning it into a plugin.
• GitHub acts as a central "source of truth" for all shared skills, organized by department or function.
• Plugins make these skills easily accessible within AI tools like Claude Code and CodeX.

Implementing and Using Plugins

• The process involves creating a GitHub repository for skills and then adding this repository as a "marketplace" within the AI tool's plugin system.
• Skills can be bundled into department-specific plugins for better organization and to avoid irrelevant access for team members.
• Enabling "auto-update" for plugins ensures that all users automatically receive the latest versions of skills.
• This solution is not limited to one AI platform and works across multiple tools like Claude and CodeX.

Advanced Features and Best Practices

• A "Skills Assistant" UI can help visualize skill connections and guide less technical team members.
• Usage tracking can identify underutilized skills for culling and highlight popular skills.
• Skills can be added to enterprise accounts as "organization plugins" for automatic distribution.
• The "skill-maxxing" philosophy emphasizes creating detailed, "fat" skills rather than overly complex agent instructions.
• A self-improvement loop can be built into skills to allow for continuous refinement based on usage and feedback.
• Creating skills for repeatable tasks is encouraged to automate online processes and build company value.
• Maintaining a personal "sandbox" repository for individual skills is recommended for backups and experimentation.

Setting up Claude Code as an AI Employee

• Workspace: A repository (repo) where the product files reside.
• Memory: Contextual information (customer needs, project goals, learnings) for Claude to understand the project.
• Brief/Plan Mode: Claude analyzes the job and proposes an approach before making changes.
• Ticket: Clear, specific assignments with defined deliverables. Avoid vague prompts.
• Eyes: Claude can inspect the application, click through flows, and identify user experience issues.
• Review: Claude can help review code against defined standards and identify must-fix, should-fix, or okay-to-ship issues.
• Schedule (Routines): Recurring tasks like daily briefs or weekly reviews that Claude can perform autonomously.
• Permissions: Defining what Claude can do freely, what it needs to ask permission for, and what remains human-owned.
• Skills, Connectors, Hooks: Creating repeatable actions (skills), integrating with other tools (connectors), and adding guardrails (hooks) to customize Claude's functionality.

Workspace Setup Example: Med Spa Lead Responder

• Organize project folders: /app, /context, /customers, /demos, /routines.
• Create essential Markdown files:
  • claude.md: Defines working style, business context, quality bar.
  • roadmap.md: Outlines current goals and priorities, specifying what's in/out of scope.
  • review.md: Contains a checklist for code review standards.
• Prompt Claude to set up the repo structure and populate these files with business context (product, buyer, pain, promise, goal).

Utilizing Plan Mode and Tickets

• Use Plan Mode for meaningful tasks by prompting Claude to inspect the app, identify files to change, outline the implementation, user experience, risks, and verification steps.
• Approve or revise Claude's plan before execution.
• Issue clear, small tickets (e.g., "Add a weightlist form with name, email, company fields and a success message") to ensure Claude knows the expected outcome.
• Implement changes in focused, reviewable commits.

Leveraging Claude's "Eyes" and Review Capabilities

• Prompt Claude to inspect the live application from a customer's perspective, checking for clarity, trust signals, and functionality.
• Utilize Claude's review capabilities, guided by review.md, to categorize issues (must-fix, should-fix, okay-to-ship).
• Consider "ultra review" for high-risk changes before production deployment.

Implementing Schedules and Parallel Agents

• Set up routines for recurring tasks, such as a daily morning brief summarizing customer pain points and risks, or a weekly ops review.
• Enable parallel agent work by using separate Claude sessions for different tasks (e.g., bug fixing, product clarity, sales enablement) within the same project context.
• Ensure each session has a clear objective and defined handoff criteria.

Permissions and Skills for Customization

• Define "safe actions" (e.g., reading files), "ask first actions" (e.g., installing dependencies), and "human-owned actions" (e.g., production deploys).
• Develop reusable "skills" for common tasks (e.g., landing page teardown, customer notes analysis).
• Use "connectors" to give Claude access to external tools (GitHub, Slack) and "hooks" to enforce workflows (formatting, tests).

7-Day Plan to Create an AI Employee

• Day 1: Create the repo brain (.md files, folders) and define core business context.
• Day 2: Utilize Plan Mode for a small product task.
• Day 3: Build one visible improvement based on the plan.
• Day 4: Use the preview loop for testing and refinement.
• Day 5: Conduct thorough code review using diffs and Claude's review standards.
• Day 6: Share the work with potential users and gather feedback.
• Day 7: Create the first routine (e.g., morning brief) to establish a daily feedback loop.

Shifting from Management to Enablement

• The term "managing agents" is problematic; shift to enabling them.
• AI agents should be empowered to break through ceilings, not just execute tasks.
• The goal is to offload administrative tasks, mirroring human management preferences.

The "Do Smart Things" Prompt

• A powerful prompt is three simple words: "do smart things."
• This prompt leverages AI's access to all context (docs, meetings, emails, etc.).
• It allows AI to proactively identify and execute valuable tasks.

Levels of Proactivity

• Employees (and agents) can be categorized by proactivity: task completion, exceeding expectations, and proactively identifying/executing new tasks.
• "Do smart things" shifts responsibility to agents to identify and execute new, valuable tasks.
• This expands breadth and scope, not by delegating more tasks, but by giving agents more agency.

Proactive and Undefined Workflows

• Proactive automations can be trigger-based (e.g., transcriptions, social posts from videos).
• The future lies in proactive, undefined workflows where AI applies its probabilistic reasoning to tasks.
• This requires agents to understand goals, have access to tools, and know what triggers action.

Building an AI-Native Workforce

• Rethink traditional job titles; focus on AI-native roles and functions.
• Create new roles like "Chief Dreaming Officer" (Phoebe) for unconventional ideas or "Assistant" (Toby) for monitoring agent performance.
• The cost-effectiveness of AI agents allows for hiring specialized, unconventional roles.

Designing the AI Workforce Structure

• Start by understanding individual agent interactions, then proactive agents, then multi-agent collaboration.
• Develop a "mission control" view to understand agent interactions and context passing.
• Recognize that not all agents require powerful models (e.g., using Haiku/Sonnet for sub-agents).

Arbitrage Opportunities in AI

• Be in the top 1% of AI users by building and utilizing basic AI workforces.
• Use AI as a watchdog for anomaly detection and insights, not just dashboards.
• Build "factories" for AI products, creating reusable primitives for faster development (e.g., login, payments, social sharing).

The Future of SaaS and Consumer Apps

• Mediocre software will die; enterprise solutions will persist due to needs for security, maintenance, and accountability.
• Consumer apps are shifting from science to art, with creativity and marketing (video, influencers) driving success over pure code.
• Opportunities exist in creating agent-first versions of existing apps or undercutting prices.

Identifying and Solving Bottlenecks

• Focus on high-value bottlenecks, not just any bottleneck.
• Video creation remains a high-value bottleneck despite AI editing assistance.
• B2B influencer marketing is a significant bottleneck for building trust.

Research Avenues for Opportunities

• Monitor YC's application trends for insights into future AI development (18 months ahead).
• Utilize resources like Matt Van Horn's "/last30days" for industry research and idea generation.
• Listen to industry leaders (e.g., CMOs) to understand fear points and emerging needs (e.g., agent-first shopping experiences).

Embracing the Weirdness

• Lean into the unexpected behaviors of AI (e.g., emoji reactions) and use them to your advantage.
• Develop multiplayer AI workforces where humans and AI collaborate seamlessly.
• Focus on systems that uplevel performance and encourage better question-asking.

Cloudflare's AI Initiatives

• Cloudflare is launching AI agent infrastructure focused on monetizing internet resources for machine usage.
• Key features include AI Crawl Control for managing AI bot access and Pay Per Crawl for charging AI crawlers.
• The Monetization Gateway extends this to any resource behind Cloudflare (web pages, APIs, data sets) using the HTTP 402 code.
• This shifts the internet model from monetizing human attention to monetizing machine-readable resources.

The New Internet Stack and Opportunities

• The internet is shifting from messy human-readable pages to structured, agent-readable resources.
• A new stack is forming: messy internet -> structured data -> agent-readable API/tools -> payment rules -> trust & analytics.
• The core question for new businesses: What resource does an agent need badly enough to pay for?

Startup Idea 1: Niche Data Refinery

• Refine messy, fragmented, or changing data within a specific niche into clean "fuel" for AI agents.
• Example: Compiling med spa competitor pricing, reviews, and hiring trends into actionable insights.
• Initial sales strategy: Target agencies or consultants serving the niche, offering them "local market intelligence."
• Progression: Report -> Dashboard -> API -> Agent-payable resource.

Startup Idea 2: Agent Readiness for Businesses

• "SEO for the agent internet": Helping businesses make their information easily understandable, trustable, comparable, and recommendable by AI agents.
• Initial approach: Offer paid audits by testing buyer intent prompts across AI tools to reveal business blind spots (e.g., incorrect pricing, buried information).
• Solution: Create agent-readable "source of truth" documents (LLM.TXT, structured FAQs, clean docs, pricing pages).
• Recurring model: Monthly prompt re-runs to measure improvements and identify ongoing needs.

Startup Idea 3: Expert Archives to Agent Tools

• Transform creators', experts', or media companies' archives (videos, podcasts, newsletters) into specific agent-powered tools.
• Focus on a single job or workflow, e.g., a cold email critique tool based on a sales expert's archive.
• Process: Transcribe/collect content -> Tag by job, topic, example -> Build one specific workflow tool.
• Benefit: Leverages creator's existing distribution and trust. Monetization via subscriptions or licensing.

AI Marketing Agents for Customer Acquisition

• AI marketing agents can automate customer acquisition, similar to how coding agents automate software development.
• This episode focuses on building two end-to-end marketing agents to generate customers.
• The strategy involves identifying "hand-raisers" – individuals showing interest through engagement with specific content.

Agent 1: LinkedIn Engagement to Cold Outreach

• Objective: Monitor LinkedIn influencers, extract engagers, enrich contact info, and initiate outbound campaigns.
• Process:
  • Identify relevant influencers and their content on LinkedIn.
  • Use scraping tools (like Ampify) to extract users who engage with specific posts.
  • Utilize a "waterfall enrichment" process with tools like GetLeads.io, Apollo, and Origami to find emails and phone numbers.
  • Validate email addresses using tools like MillionVerifier to maintain deliverability.
  • Set up inbox infrastructure (e.g., InboxKit, Instantly AI, Hypertide) using burner domains to protect core domains.
  • Send cold emails via platforms like Instantly AI.
  • Conduct LinkedIn DM campaigns using tools like HeyReach or BotDog.
• Agent Functionality: Agents can monitor inboxes, respond to inquiries, book demos, and manage follow-ups autonomously.
• Infrastructure Cost: Estimated at $100-$200/month for sending software and inboxes.

Agent 2: Scalable Social Content Creation

• Objective: Generate consistent, high-quality social media content for teams or individuals.
• Process:
  • Source raw material from human conversations (e.g., interviews, sales calls, podcasts, internal comms).
  • Use an LLM (e.g., Claude Sonnet) via API to transform source material into social media posts.
  • Schedule posts across multiple LinkedIn accounts using tools like Ordinal.
  • Feed analytics data back to the agent to identify high-performing content and remix successful ideas.
• Benefit: Allows a single person to manage numerous social media accounts, leveraging earned media and platform payments.
• Alternative Strategy: Create theme-based or topic-based pages (e.g., @GrowthTactics) instead of personal brands.

Core Concepts & Tools

• Marketing Agents: Software with potential thinking loops and live data streams performing specific jobs.
• Tools Mentioned: Codeex, Claude Code, Ampify, API Maestro, Cloud Code, GetLeads.io, Apollo.io, Origami, MillionVerifier, InboxKit, Instantly AI, Hypertide, HeyReach, BotDog, Ordinal.
• Infrastructure: Requires a server (e.g., Railway) for deploying agents and a data pipeline/warehouse (e.g., ClickHouse).
• LLM Usage: Emphasizes using LLMs for specific tasks rather than paying per token for every action; focus on building software that utilizes compute.
• Data Sourcing: Leveraging AI to query internal data sources (Slack, Notion, Gong transcripts) for content ideas.

What is Graph Engineering?

• Graph engineering designs workflows around AI to manage complex tasks beyond a single chat.
• It transforms messy AI tasks into manageable, structured workflows with defined steps, checks, and approvals.
• It's a progression from prompt engineering (better questions) and context engineering (better information).

Core Concepts

• Graph: Jobs connected by arrows, representing workflow steps and the flow of information (state).
• Knowledge Graph: Helps AI reason over relationships within data (e.g., customer-company, product-tool).
• Agent Graph: Focuses on how work moves through a workflow (e.g., planner, researchers, skeptic, merger, approver). This episode primarily covers agent graphs.

When to Use Graph Engineering

• For tasks with multiple steps, sources, paths, checks, or approvals.
• Examples: Deep research, go-to-market plans, support triage, code review, sales prep, feedback synthesis, recurring content.
• Not ideal for simple tasks like brainstorming 10 names or summarizing a short email.

The Graph Workflow Example (Startup Idea Validation)

• Initial Question: "Should I launch an AI bookkeeping product for Shopify merchants?"
• Graph Steps:
• Planner: Defines necessary research angles (customer pain, competitors, distribution, pricing, risk).
• Parallel Researchers: Investigate customer pain, competitors, and distribution channels independently.
• Skeptic: Critically evaluates findings for unsupported claims, stale data, or overlooked competitors.
• Merger: Synthesizes surviving evidence into a recommendation (pursue, pause, kill, next steps).
• Human Approval: Final decision-maker reviews the recommendation before action.

Implementation Levels

• Level 1 (Manual): Draw the graph in a tool like Excalidraw, assign jobs to different lanes, and run them manually. Focus on structure, not automation.
• Level 2 (File-Based): Each step writes output to separate files (e.g., `customer_research.md`), creating a paper trail.
• Level 3 (Automated): Use tools like LangGraph, AutoGen, GraphFlow, or Make.com for orchestration.

Benefits and Cautions

• Improves quality, consistency, and delegation.
• Turns AI work into an "operating system" with memory and compounding value.
• Caution: More agents don't automatically mean better output; aim for the smallest graph that improves quality.
• A good graph removes "fake waiting," separates workers from checkers, and places human approval strategically.

Introduction & Drinks

• The host begins by addressing the sentiment that "software is dead" and the decline of indie hacking.
• He engages the chat to decide which energy drink to have: Red Bull Juneberry or Monster White Zero Ultra.
• The chat overwhelmingly votes for Monster, which the host tries for the first time, describing it as "super sweet" and tasting like Mountain Dew.

The "Software is Dead" Argument

• The host discusses tweets from John Yongfuk and Levels IO, who report declining signups, conversions, traffic, and revenue.
• Key reasons cited are decreased Google referrals and AI cannibalizing existing app markets.
• The consensus in many replies is that starting a software company in 2026 is futile, with AI dominating.

Counterarguments & New Opportunities

• The host argues against the "software is dead" narrative, emphasizing adaptation.
• He highlights that core businesses like CRMs and payment systems are not going away.
• New opportunities lie in super niche products, creator-led businesses, strong meta ad skills, data-first products, and maintaining software others won't.
• The primary insight is that code is no longer the unfair advantage; distribution, niche focus, data, and network effects are.
• AI-first product development and agentic-first design are key areas for future innovation.
• Hardware combined with AI is also presented as a significant opportunity.

Adapting the "Indie Hacker" Model

• The traditional "indie hacker" model of coding a tool and selling it for a low monthly fee is becoming less viable due to increased competition and changing traffic sources (away from Google SEO).
• The emphasis needs to shift from being a developer to becoming a marketer and distribution expert.
• Community and organic marketing are crucial moats.
• The host suggests focusing on earning attention and crafting affinity through content.

The Future Landscape

• The demand for software is expected to increase as AI expands the total addressable market.
• Entrepreneurship core remains solving problems and driving value; AI creates new problems to solve.
• Coding skills are still helpful for reducing friction, but marketing and problem-solving are paramount.
• New opportunities exist in agentic-first products, AI-powered services, and niche markets.
• The sentiment around AI is often negative outside tech circles; focus should be on the outcome and value provided, not just "AI-powered."

Key Takeaways & Conclusion

• The "unfair advantage" has shifted from code to distribution, data, network effects, and creativity.
• Building a business is the focus, not the label "indie hacker," "creator economy," or "SAS."
• The core is still solving problems and driving value; AI has created new problems and accelerated solutions.
• The future belongs to those who are creative, craft stories, understand branding, and connect with customers.
• The host concludes that now is an exciting time to build, not a time to despair.

Buzz Overview

• Buzz is described as an "agentic version of Slack" where teammates are AI agents.
• It aims to allow software building on the fly and is seen as a glimpse into the future of work.
• Buzz is built on the open protocol Nostr.

Key Features and Benefits

• Agents as First-Class Citizens: Unlike Slack integrations, agents are integral team members.
• Swappable Harnesses: Users can switch AI models (e.g., Cloud Code, Codex, Goose) without losing chat context. This addresses "model fatigue."
• Audio Huddles with Agents: Allows for live voice conversations with agents, fostering more natural interaction for creative tasks.
• Integrated Git and Projects View: Agents can create projects, feature branches, and work in parallel on copies of code, with a dedicated projects view. Buzz also has its own hosted Git hosting on "relays."
• Automated Software Deployment: Agents can build and deploy applications (e.g., CRM, tweet leaderboard) to platforms like Railway, providing screenshots and links.
• Openness and Integration: Built on open protocols, allowing for easier integration compared to proprietary platforms like Slack.
• Context Engine: Buzz prioritizes context, enabling agents to leverage past conversations and data for complex tasks.
• Bitcoin Lightning Integration (Potential): Due to Nostr's connection with Bitcoin Lightning, future integration for payments (compute, tips) is anticipated.
• Shared Compute: Allows users to share local LLM models running on a single machine, benefiting small teams or individuals with limited resources.

Use Cases and Target Audience

• Ideal for solopreneurs and small teams for brainstorming, product building, and ideation.
• Can automate "boring" business tasks, creating an unfair advantage.
• Useful for generating proposals, marketing content, and analyzing data.
• Potential for community engagement with public/private channels.

Tips and Considerations

• Agent Configuration: Pin agents to specific models (e.g., Fable, Sonnet) to manage token usage. Consider a "Chief Agent Officer" for delegating tasks.
• Skills: Globally installed skills from other platforms (like Cloud Code) are accessible to Buzz agents.
• Early Stage: Buzz is beta/alpha software; some features might be slow or not fully polished.
• Open Source Advantage: Being open source on an open protocol allows for community improvements and rapid iteration.

What is a Marketing Agent?

• A marketing agent needs to solve the data problem for unified clarity across the entire pipeline.
• It must autonomously make decisions with a thinking loop.
• Infrastructure requires a data pipeline, data warehouse, and cloud hosting for agents.
• Agents automate processes and improve based on data feedback, not necessarily AGI.
• An agent can now autonomously run Facebook ads: researching pain points, creating on-brand static and AI avatar UGC, publishing, optimizing, and iterating.

Startup Idea: AI for WordPress

• WordPress powers over 40% of the internet, offering a blue ocean market.
• Bundle essential plugins (forms, CRM, etc.) into a single package sold via tokens (e.g., $29/month).
• Opportunity to build AI-first versions of existing popular WordPress plugins:
  • Yoast SEO: Agent writes meta descriptions, restructures content, adds internal links.
  • WP Forms: Conversational AI agent to qualify leads and answer questions.
  • WooCommerce: AI storekeeper for product descriptions and abandoned cart flows.
  • Kismmet: AI spam and security prevention.
• Bundling offers cost savings and enhanced functionality compared to multiple individual plugins.

Deploying Marketing Agents for Facebook Ads

• Facebook's Andromeda Algorithm: Focus creative on pain points and outcomes, not just interest-based targeting.
• Creative Process:
  • Research Pain Points: Use tools like Perplexity to scrape Reddit for real user complaints and desired outcomes.
  • Identify Key Issues: Focus ads on pain points like plugin complexity, site performance/speed, and security/maintenance.
  • Generate Creative: Use AI tools like Kai AI (for images/video) and HeyGen/Seed Dance (for video avatars) to create on-brand assets.
  • Brand Style Guide Compliance: Use vision models to ensure ads adhere to brand fonts and colors.
  • AI Avatar UGC: Create video content where avatars discuss identified pain points.
• Facebook Marketing API: Use only for publishing, turning off, or promoting ads, not for excessive data pulling, to avoid account bans.

Infrastructure for Marketing Agents

• Data Pipeline & Warehouse:
  • Data Pipeline: Use open-source tools like Airbyte to pipe data from various sources.
  • Data Warehouse: Use ClickHouse to unify data from Facebook Ads, Google Analytics, PostHog, CRM, Stripe, etc.
  • Purpose: Allows agents to understand all data in context, connecting ad spend to revenue.
• Agent Hosting: Deploy agents as code in any cloud environment (Heroku, Railway, etc.).
• Agent Function: A decision tree that uses live data streams and LLM thinking to optimize for revenue.

Operationalizing Marketing Agents

• Autonomous Ad Management: Agents can automatically upload ads, monitor performance, turn off losers, and allocate budget to winners.
• Solving Entropy (Agent Staleness):
  • Scrape competitor ads from the Facebook Ads Library.
  • Extract insights from YouTube and podcast transcripts (e.g., WordPress channels).
  • Use tools like Viral Low to identify trending content formats and themes from TikTok/Reels.
• Virtual Employee Concept: Agents act as specialized virtual employees focused on specific marketing channels.
• Evolving Marketer Role: Marketers become "agent jockeys," leveraging domain knowledge to build and manage these systems.
• Speed of Deployment: Founders can set up a comprehensive ad system in under two hours.

Future Agent Capabilities

• Google Ads agents for full account management.
• Influencer outreach agents for research, cold emailing, and negotiation.
• Cold email agents with inbox management for response and booking.
• TikTok/Reels agents for scaled content creation and posting.
• SEO agents for keyword research, article writing, and on-brand content.
• AI search agents for citation building and outreach.
• Social media management agents for LinkedIn, Twitter, etc.
• Email newsletter agents leveraging podcast content and lead magnets.

AI Agents: The Most Valuable Skill

• Managing teams of AI agents is the most valuable skill for outperforming others in the AI age.
• This skill enables individuals to become "agent pros," regardless of their background (founder, solopreneur, parent, etc.).
• Becoming a world-class agent operator means acting as a manager of AI agents.

Managing AI Agents: Setup and Tools

• Ryan Carson's Setup: Uses a 52-inch monitor with 8 screens for multitasking agents, a vertical mouse for wrist pain, and a button connected to Whisper Flow.
• Security: Emphasizes keeping API keys secure in a password manager and never giving production keys directly to agents. Agents should request keys when needed, with manual input from the user.
• Devon AI: Highlighted as a powerful, albeit expensive, software factory tool that works well with cloud agents.
• Local vs. Cloud Development: Argues strongly for cloud development (using VMs in the cloud) over local development, stating local development significantly limits output and efficiency. Cloud environments eliminate code collision issues and mental overhead.

The Power of Cloud Agents

• Cloud agents, running in virtual machines (VMs) in the cloud, allow for infinite, simultaneous development environments without code conflicts.
• This approach drastically increases output and efficiency, making local development seem outdated ("caveman" approach).
• Even front-end development, after initial wireframing, should ideally move to the cloud.

Managing High-Stakes Decisions and Workflow

• The nature of work shifts to making high-stakes decisions frequently throughout the day.
• Organization: Pin important threads/tasks and separate them from minor ones.
• Cadence: Establish a rhythm for checking agents and making decisions (e.g., every 25 minutes) to avoid mental exhaustion.
• High Output: Expect and manage significantly higher output, with examples of shipping 20-40+ Pull Requests per day.
• Phone Usage: A significant portion of work, including high-stakes decisions, can and should be done from a phone.

Automations for Efficiency

• QA Bug Testing: Automate user experience flows (like sign-up and onboarding) to run regularly and identify bugs. Devon's "playbooks" are mentioned for defining these processes.
• Self-Improvement Loop: Automate the grading of agent conversations (e.g., a paralegal agent) and trigger fixes for any low-scoring interactions.
• Production Watchdog: Daily automation that summarizes activity for paid customers, providing a "chief of staff" overview of important events and potential issues.
• Browser Testing: Crucial for agents to accurately test in browsers, with Devon recording, annotating, and self-fixing based on video analysis.

Cost Management and Model Routing

• While token costs can be high ($20k/month mentioned), using the right model for the right task is key.
• Independent agent labs (Devon, AMP, Factory, Cursor) are incentivized to optimize for affordability and model routing (e.g., using cheaper, fine-tuned models like SUI 1.7 for specific tasks).
• Avoid building a company's software factory solely on frontier lab stacks (like Cloud Code or CodeX) due to potential vendor lock-in and lack of long-term cost optimization.
• Building a "software factory" is essential as companies scale, and independent labs provide this infrastructure.
• Using CodeX can be beneficial for "free" token usage for personal information work, but not for core company infrastructure.

Building Credibility and Sharing Knowledge

• Share learnings publicly, especially on platforms like X (formerly Twitter), to build credibility and network.
• Writing helpful articles and sharing knowledge, even if not perfectly polished, opens doors and fosters relationships.
• Learning and doing are paramount; "there is no try, only do."

Naval's Calendar Philosophy

• Naval has deleted his calendar and refuses to schedule specific times for interactions, inspired by advice from Marc Andreessen and Jack Dorsey.
• He operates via text and has a "hostile email autoresponder" to signal unavailability.
• The hosts suggest this is a luxury of success and not universally applicable advice.
• The core takeaway is that less scheduled time can create space for spontaneity and saying "yes" to unexpected opportunities.

AI and Business Growth

• To go "AI native," companies must make their data legible to AI (meeting notes, SOPs).
• Then, map "jobs to be done" for each role and build AI agents to assist or automate tasks, creating loops.
• For new companies without audiences, customer acquisition involves paid ads or organic methods like SEO and content creation.
• Content creation (podcasts, YouTube, X posts) can be a slow but effective way to funnel targeted customers.
• The challenge in crowded markets (like indie games) is discoverability, requiring original, engaging content and community building.
• The "devil's advocate" stance against AI's profitability is challenged; many businesses are quietly using AI to generate revenue and build competitive moats through brand and distribution.

Productivity vs. Leverage

• Obsessing over productivity can be a form of procrastination; focus should be on high-leverage activities.
• The question "If I could only work 4 hours a week, what would I do?" helps identify key business priorities.
• True success often comes from finding leverage, not just optimizing personal output.
• Productivity is personal; what works for one person may not work for another.

Content Strategy and Branding

• Effective marketing involves turning strangers into customers through paid ads or organic means.
• For non-US builders, starting content in a local language (e.g., French) is viable, with a potential expansion to English later.
• Platforms offer translation tools, but comfort in native language is a strong starting point.
• Story and brand are crucial in a world where anyone can build anything; a compelling narrative drives interest.
• The "why now" for a product or game is essential for market entry.

Entrepreneurial Origins

• Personal origin stories in business are often less glamorous than portrayed, involving practical needs like making money and following opportunities.
• Riding market waves (like the initial demand for UX designers) can be a successful strategy.
• Building internet businesses can feel like a video game with a focus on growth metrics ("number go up").

What is a Forward Deployed Engineer (FDE)?

• FDEs bridge the gap between general AI intelligence and a company's specific context.
• They determine where and how AI intelligence should be applied within a business.
• The role requires a blend of deep technical understanding and strong communication skills.

The FDE Role and Its Evolution

• Palantir popularized the FDE role by deploying engineers on-site to customize their software stack for clients.
• The AI age demands customized agents for every company, driving the need for FDEs.
• The advantage in AI is no longer just having access to intelligence, but in how effectively it's deployed.

Key Stages of an FDE's Work

• Understanding Business Reality: Deeply learning current workflows, processes, and exceptions through observation and interviews. This is the most time-consuming part.
• FDE Judgment: Deciding where AI should be applied, balancing risk and ROI. Not every workflow needs AI; FDEs identify tasks best suited for LLMs and those for deterministic software.
• Building the Deployed AI System: This can range from configuring workflows on platforms to writing production code, depending on the company and role.

Skills and Compensation

• FDEs need to be excellent communicators and adept at understanding business needs, alongside technical proficiency.
• Salaries can range from $150,000 base with equity to $1 million per year.
• The role requires both consulting-like business acumen and software engineering capabilities.

The FDE Process: Audit, Eval, Deploy

• Audit: Analyze existing workflows, identify bottlenecks, and map processes. This is crucial for understanding the "real" process, not just the documented one.
• Eval (Evaluation): Develop metrics and testing suites to ensure the AI system performs correctly, especially for non-deterministic tasks. Human-in-the-loop feedback is vital.
• Deployment: Integrate AI solutions into existing systems, focusing on de-risking the adoption for the client and demonstrating measurable value (revenue uplift, risk mitigation, cost savings).

A 30-Day Plan to Become an FDE

• Week 1: Build a Looping Agent - Create a functional agent for a real enterprise workflow, focusing on agent looping, tool usage, guardrails, context, memory, and audit trails.
• Week 2: System Recovery - Engineer the agent to handle failure modes and exceptions effectively, building for unhappy paths.
• Week 3: Measurable Viability - Optimize the agent, use cheaper models for subtasks, implement retry logic, and measure economic impact across revenue, risk, and cost.
• Week 4: Defend the System - Rehearse the technical architecture and business value proposition, pitching the agent to businesses to gain feedback and refine understanding.

Customer Acquisition Strategies

• Start by building an audience and community (social media, paid/free communities, IRL events) before creating a product or service.
• Focus on creating a media company first, then build a product.
• Organic growth through audience building is the hard but rewarding way.
• Advertising (e.g., Facebook Ads) is the easier but more expensive option.
• For individuals with full-time jobs, dedicate specific days (e.g., Friday-Sunday) to building a business.

Sales and Entrepreneurship Mindset

• In the age of the generalist, embrace learning multiple skills, including sales.
• The owner's primary role is promotion and sales, not just delivery.
• If you can't sell, find a co-founder who can, but be prepared to give up equity.
• Don't be afraid to make cold calls and get rejected; learning comes from doing.

Navigating the AI Age

• Ignoring AI is a bad idea; conversely, trying to use every new tool can lead to overstimulation.
• The sweet spot is using established AI tools and selectively exploring new ones.
• Build things agents want, not just things humans want.
• AI is creating new opportunities by making previously expensive tasks (like creating marketing assets) accessible to individuals.
• Design agencies need to shift from pure execution to strategy and responsibility-bearing.
• Small businesses often prefer to outsource tasks like website development even if they have the capability to do it themselves, valuing convenience and responsibility.

Timely Insights and Trends

• Sega originally stood for "Service Games" and supplied machines to US military bases.
• Success in business, especially for smaller ventures, doesn't always require strict goal-setting; following curiosity and experimenting is viable.
• Cloudflare's new monetization gateway could enable charging AI agents for web page access, creating opportunities for agent-focused content.
• AI agents may soon have wallets and email addresses, enabling transactions and a new internet economy.
• Reading books offers a significant competitive advantage in terms of information retention and critical thinking.
• Local LLMs offer data privacy benefits but require significant compute power and investment.
• A hybrid approach using both powerful frontier models and smaller, local/open-source models is likely the future.

AI Business Model Overview

• Core Offer: AI Tools Assessment for small businesses ($999).
• Value Proposition: Prescribe 3-7 off-the-shelf AI tools to save 5-10 hours/week. 100% money-back guarantee if less than 5 hours saved.
• Target Clients: Small business owners (2-20 employees, $500K-$5M revenue).
• No Prerequisites: No audience, capital, or coding skills needed. Understanding AI is key.

Fulfilling the AI Tools Assessment (4 Phases)

• Phase 1: Discovery Call: Record a 45-min structured interview to identify pain points and time drains.
• Phase 2: AI Analysis: Use AI (like Claude) to analyze the transcript, identify pain points, and research relevant off-the-shelf AI/SaaS tools. Quality assurance is crucial.
• Phase 3: Report Generation: Create a client-facing report (using tools like Claw Design) detailing executive summary, effort vs. impact matrix, recommended tools, a 4-day quick start plan, and financial impact.
• Phase 4: Review Call: A 30-min call to screen-share the report, walk through recommendations, and ask closing questions about implementation preference and timeline.

Upsell Opportunities & Recurring Revenue

• Process Redesign: Fixing broken processes (e.g., $3K-$3.5K).
• Automation Builds: Simple workflows using Zapier, Make.com, etc. (e.g., $1.5K).
• Knowledge Systems: Building custom GPTs trained on specific business data.
• Custom Workflows: Building proprietary processes, often using Claude skills. Potential for recurring revenue.
• Full Implementation: Bundling multiple services.
• AI Concierge (AI Done-With-You): Monthly retainer ($1.2K-$2K+) for two 45-min calls/month focusing on building Claude skills. Includes unlimited Voxer access (12-hour response time).

Client Acquisition Strategies (No Capital/Audience)

• Host Local AI Meetups: Position as an expert, network, and follow up with attendees.
• Door Knocking: Direct sales to local businesses for immediate results.
• LinkedIn DMs: Targeted, pain-point-focused outreach (not pitching in the first message).
• Free Audits/Assessments: Offer mini-audits to your existing network.
• Agency Partnerships: Collaborate with accountants, insurance agents, etc., for referrals.
• AI Office Hours: Offer free AI consulting at co-working spaces.
• Post Your Wins: Document and share small successes publicly.

Niche Strategy & Long-Term Growth

• Specialization: Focus on a specific industry vertical or geographic area to stand out.
• Unique Selling Proposition: Develop a personal brand as the go-to expert for a niche.
• Tiered Pricing: Gradually increase prices and offer different package tiers to create urgency.
• Continuous Upselling: Identify further needs and opportunities within existing client relationships.

What is Loop Engineering?

• Loop engineering involves creating automated cycles, similar to "build, measure, learn," but for AI agents.
• The concept has gained popularity, with a vision of software self-building and achieving product-market fit.
• It's an extension of existing concepts like Lean Startup and continuous improvement.

Applying Loops to Business

• SEO Improvement:
• An AI agent can continuously monitor Google rankings for target keywords.
• It can analyze ranking data from Google Search Console and Data for SEO to identify areas for improvement.
• The agent can make changes (e.g., update meta tags, optimize content) and then re-evaluate rankings, iterating over time.
• This loop can run monthly or bi-weekly, aiming to improve search engine positions and drive traffic.
• Facebook Ads:
• AI agents can generate ad variants, test them, and allocate more budget to successful ones.
• This automates the process of ad optimization, which is often done by human agencies.
• It can be a mix of human-generated copy with AI optimization for volume and testing.
• Product Feedback Loop (The "Holy Grail"):
• An advanced loop where an AI agent reads customer feedback, analytics, and logs.
• It identifies pain points, prototypes features, and fixes bugs based on this data.
• The agent can prioritize based on chosen KPIs like DAU, retention, or NPS, or even let the AI decide.
• This could potentially lead to a self-building business, though it carries risk.

Implementation and Cost

• Loops can be implemented using tools like Claude Code or Codhex.
• Connecting AI agents to real data sources (Google Search Console, etc.) is crucial.
• The cost is generally low, especially compared to hiring agencies, with runs potentially costing less than $5 in tokens.
• For those on max plans, token cost is less of a concern; on lower plans, using open-source models might be advisable.
• Automation tools (routines, automations) can schedule these loops to run periodically.

Grok 4.5 as an AI Co-founder

• Grok 4.5 is presented as a significant upgrade, enabling AI agents like Hermes to function as AI co-founders.
• The key unlock is giving AI agents access to various tools, capabilities, and context, transforming them from automation tools into genuine collaborators.
• Grok 4.5 offers intelligence comparable to top-tier models but at a fraction of the cost and with significantly increased speed.

Setting Up and Using AI Agents

• The setup involves using platforms like Orgo to host AI agents on cloud computers, giving them access to tools, email, phone numbers, and even debit cards.
• Users are encouraged not to be hesitant about granting agents extensive tool access, as this unlocks their full potential.
• Demonstrations include an agent spinning up new computers, accessing dashboards, and opening applications like Chrome, showcasing its operational capabilities.

Performance and Cost-Effectiveness

• Grok 4.5 is highlighted for its speed and affordability, significantly outperforming previous models in task completion time and cost efficiency.
• While faster and cheaper, the increased productivity means users may end up spending more due to the ability to accomplish more tasks.
• The model is described as a "cheap Ferrari" compared to previous "expensive Toyotas," offering a dramatic improvement in performance and value.

Tool Integration and Use Cases

• The presentation emphasizes integrating various tools such as X, Idea Browser MCP, Composio, Agent Mail, Agent Phone, Vid IQ, and Latitude for enhanced agent functionality.
• Use cases demonstrated include generating startup ideas, creating landing pages, crafting marketing content (thumbnails, scripts), and building email outreach campaigns.
• The ability to seamlessly generate content, from website copy to YouTube thumbnails and outreach sequences, showcases the agent's comprehensive capabilities.

The Future of AI Agents

• The trend towards agents and AI employees is likened to the rise of marketing agencies in the early 2010s, representing a significant opportunity.
• The focus is shifting from traditional interfaces to conversational interactions with AI agents via text or voice.
• The shrinking gap between idea conception and implementation is a key takeaway, empowering users to bring their ambitious projects to reality quickly with AI.

Baby Food App

• Take: A baby food app is making $1 million/month by providing first 100 food ideas, meal plans, and tracking reactions.
• Analysis: Parents are willing to trust and pay more for expensive baby apps due to the perceived safety and importance of their child's health.
• Opportunities: Niche down into specific dietary needs (e.g., allergies) or pet food equivalents.

Customer Support & Engineering

• Take: Customer support is increasingly "eating" engineering as AI enables non-engineers to make product changes.
• Analysis: AI and LLMs can process customer support data to identify issues and prototype solutions, creating feedback loops that inform engineering.
• Counterpoint: Highlight human-only customer support as a premium differentiator against AI-driven services.

Team Size in AI Era

• Take: The "two-pizza team" concept is outdated; AI enables smaller, more efficient teams.
• Analysis: AI agents can handle tasks previously requiring multiple people, suggesting a "one-pizza" team might suffice, especially when paired with strong sales/marketing.
• Insight: Focus on creating AI agents to solve specific "jobs to be done" before hiring.

Decline of Social Spaces

• Take: There's a societal trend of fewer places to relax and socialize.
• Opportunity: Significant opportunity to create "third spaces" that combat loneliness, but these require capital investment.
• Approach: Leverage digital-first strategies (audience building, pre-sales) to fund physical spaces and create unique, community-focused experiences.

AI Marketing vs. Product Development

• Take: Bragging about shipping AI software is like bragging about taking many photos; the focus should be on marketing and actual value.
• Critique: CEOs are getting lost in building AI systems instead of focusing on revenue and core business functions.
• Analogy: AI's current arbitrage opportunity is compared to early Facebook Ads, offering significant ROI for those who leverage it effectively.

Systems vs. Action

• Take: Building systems is a luxury, not a starting point for entrepreneurs.
• Argument: Focus on revenue and product-market fit first; build systems only when things are breaking.
• Emphasis: Prioritize marketing and sales over premature systemization, especially in the early stages of a business.

Specialized AI Agents

• Take: Managing one AI agent with hundreds of skills is a nightmare; specialized, role-based sub-agents are needed.
• Reasoning: Specialization limits the "blast radius" when things go wrong and improves manageability.
• Example: Separating an SEO agent from a reporting agent prevents issues in one from breaking the other.

AI Pace and Adoption

• Take: The rapid pace of AI development is overwhelming, but there's a balanced approach to staying informed.
• Analogy: The current AI landscape is akin to the early days of Facebook Ads, offering arbitrage opportunities.
• Advice: Focus on a few key updates weekly, rather than succumbing to FOMO or trying to master everything.

Wealth & Leisure

• Take: True wealth is being able to spend 52 days studying a Slinky without financial worry.
• Inspiration: PewDiePie exemplifies a European approach to wealth: achieving financial security and then pursuing personal interests.

Marketing Stunts vs. Product

• Take: Companies spending more on marketing stunts than product development are a red flag.
• Nuance: While big marketing pushes can be suspicious for startups, established companies (like Sony) can afford them.
• Core Idea: Focus on organic growth and product-market fit before large-scale marketing stunts.

Seniors & Tech

• Opportunity: Connect young people with seniors for companionship and AI education.
• Market Insight: Older adults don't always see themselves as "old" and can be underserved by tech marketing that targets them as such.

Introduction to Codeex and GPT 5.6

• Codeex is OpenAI's answer to cloud coding and knowledge work, presented as a more streamlined and powerful alternative to previous attempts.
• GPT 5.6 is highlighted as the most usable, powerful, and fastest model for knowledge work and coding, not as a "tactical nuke" like Fable, but as a highly capable tool for daily tasks.

Codeex for Personal and Business Life

• Codeex can be used as an "operating system for work," managing emails, writing, and even training models.
• A demo shows an app that processes emails, summarizes content, and drafts replies, functioning similarly to how executives managed correspondence in the past.
• Codeex can manage company feeds, processing unread Slacks and meeting notes into actionable cards, learning user preferences over time.
• The system revises prompts based on user feedback (archiving, liking, replying), creating a personalized workflow.

Key Features and Capabilities

• GPT 5.6 offers improved email response drafting, making fewer mistakes and sounding more natural.
• It seamlessly connects to Slack and browses the web, assisting with tasks like apartment searching and visualizing room layouts.
• Codeex allows threads to message each other, enabling the orchestration of multiple agents.
• The in-app browser is crucial, allowing users to interact with any SaaS app, with agents providing context and assistance.
• The "SAS apocalypse" is seen as a misnomer; SaaS apps designed for agent integration within Codeex are highly valuable.

Building with Codeex and "Turnaround" Example

• The video demonstrates building a SaaS app called "Turnaround" to showcase active maintenance of software, displaying metrics like commits and response times.
• GPT 5.6 is described as "A tier," providing good results but requiring more user input than the "S+ tier" Fable for complex tasks.
• The "LFG" (Loop, Plan, Execute, Review, Compound) and "Goal" (long-term objective setting) features from a compound engineering plugin are used.
• Codeex's design capabilities have improved, though it may default to preferences like "warm paper backgrounds."
• The concept of "agent-native" SaaS apps, designed for co-collaboration between users and agents within Codeex, is emphasized as a major opportunity.
• The development of app stores for Codeex-native applications is anticipated, mirroring the success of mobile app stores.
• The "Turnaround" demo reached about 70% completion, with the speaker suggesting using Claude for better design refinement.

Pirates vs. Architects and Workflow Integration

• A distinction is made between "pirates" (who build functional 70% solutions) and "architects" (who polish and refine systems).
• The speaker identifies as a "pirate" and partners with "architects" to bring ideas to completion.
• Other Codeex uses include a "mail room" for routing emails to agents and "router threads" for managing information flow.
• Users are encouraged to download Codeex, grant it computer access, and use its "Chronicle" feature (optional, local data) for context.
• The "Record and Replay" plugin allows users to demonstrate tasks, which Codeex then turns into a repeatable skill.
• The core advice is to focus on automating repetitive tasks and shifting the mindset from doing work to managing the system that does the work.
• Starting with simple, impactful automations rather than complex, overwhelming setups is recommended.
• Getting into new AI tools should be driven by curiosity and genuine needs, not FOMO.

Agents vs. SaaS

• Agents are the new SaaS, selling "work" instead of just "tools."
• The total addressable market for agents is human capital, making it larger than SaaS.

The Product is the Job

• Focus on agents that take over a job a team no longer has to do manually.
• Examples: AI superhost for restaurants (Slang AI), AI dispatchers for home services (Same Day).
• Mental model: Agent handles an annoying job better, faster, and cheaper than manual labor.

Pick a Workflow with a Paycheck

• Identify workflows where people are already paying for the work (employees, agencies).
• A good agent workflow is frequent, has a clear finish line, touches existing software, has learnable edge cases, and causes noticeable pain when done poorly.
• Steps to find an idea: Pick a niche, list 20 jobs people complain about, score them on frequency, pain cost, ease of completion, software access, and budget ownership.

Shadow the Human & Spec the Agent

• Before building, shadow a human performing the job to understand nuances and edge cases.
• An agent spec should include: what wakes it up, context needed, tools, allowed actions, approval needs, escalation points, and success metrics.

Build the Minimal Useful Agent (MUA)

• Start small with one of four types:
• Draft and Approve: Agent drafts a response, human approves.
• Triage: Classifies and routes inbound work.
• Coordinator: Manages tasks between systems and people.
• Bounded Action: Performs specific tasks under clear rules.

The Product Wrapper Creates Trust

• The agent performs the work, but the wrapper provides logs, approvals, controls, and testing for trust.
• "Control rooms" (dashboards) show outcomes, handoffs, and analytics.
• Use "evals" (test sets of real examples) to measure agent performance and build trust.

Sell the Pilot like Labor, Then Productize

• Start with 2-3 pilot customers in the same niche and workflow, selling the outcome.
• Charge a setup fee and monthly fee, eventually moving to usage or outcome-based pricing.
• Productize by identifying repeatable patterns from successful pilots (e.g., common scripts, follow-up processes).

Distribution: Workflow Teardowns

• Create content (e.g., videos, memes) contrasting the old, painful way of doing a job with the new agent-assisted way.
• Focus on one platform and build an audience.

30-Day Agent Business Plan

• Day 1-4: Research niche, interview operators, define workflow, write agent spec.
• Day 5-6: Manually test AI, build MUA.
• Day 7: Create eval set.
• Week 2: Sell 2 pilots.
• Week 3: Build product wrapper.
• Week 4: Publish workflow teardowns, turn pilots into proof, refine content strategy.

AI Agent Management

• Skill: Setting up, managing, and running local AI models, a more advanced form of prompt engineering.
• Value: Companies will need individuals to integrate multiple AI tools into an operating system, creating specialized agents (customer support, research, sales).
• Local AI: Important for privacy, cost, latency, and control, using tools like Olama or LM Studio.
• Learning Rep: Build a daily briefing agent for yourself with specific sources, calendar, notes, and links, including a rule for approval before sending.

Distribution Marketing

• Skill: Understanding where attention already exists, people's anxieties, and their language to build trust before selling.
• Value: With easy product creation, building demand and making people care is crucial. Marketers become researchers, storytellers, media operators, and community builders.
• Learning Rep: Map a niche's attention sources (newsletters, creators, forums, podcasts), identify a core pain point sentence, and write 20 hooks for it, focusing on existing desires.

Robotics Engineering & Manufacturing

• Skill: Building hardware, integrating AI, and sourcing manufacturing.
• Value: The next decade rewards moving atoms, not just pixels. Open-source projects and low-cost components make robotics more accessible.
• Learning Rep: Assemble a low-cost robot arm with a camera, teach it a simple task (e.g., sorting objects), and document failures. Learn basics of working with suppliers on platforms like Alibaba.

Curation & Short-Form Video

• Skill: Making sense of information in public, translating new developments for a niche, and creating authentic short-form content.
• Value: With abundant AI-generated content, human curators who filter, explain, and provide takes are valuable.
• Learning Rep: Conduct a 7-day curation sprint: pick a niche, find three relevant items daily, and create a short video using the structure "I saw this. Most people think it means X. I think it means Y. Here's the move."

Builder Distributor

• Skill: Simultaneously shipping products and driving distribution.
• Value: This is crucial for founders, enabling one person to prototype, market, launch, gather feedback, and iterate without handoffs.
• Learning Rep: Complete a "48-hour loop": build the smallest version of a solution for a personal problem using AI, then create 10 distribution pieces (video, posts, DMs, landing page).

IRL Community Building

• Skill: Creating valuable in-person experiences that foster belonging, trust, and connection.
• Value: As work becomes more digital, real-world interactions and curated networks become scarce and valuable.
• Learning Rep: Host small gatherings (dinner, walk) around a specific question, invite relevant people, and send a recap that turns the room into a network.

Introduction to GLM 5.2

• GLM 5.2 is a new open-source local AI model generating buzz for its capabilities.
• It aims to be a "ChatGPT moment" for local AI, offering competitive performance.

GLM 5.2 Performance and Benchmarks

• Features a 1 million token context window and scores well on benchmarks (e.g., 81 points on Terminal Bench 2.1).
• Competes closely with models like Opus 4.8, performing well on long-horizon tasks.
• While benchmarks are informative, practical testing and building are key to understanding performance.

Setting Up GLM 5.2

• Can be run locally if hardware permits, or via cloud providers like OpenRouter.
• Cursor Setup: Get an API key from Z.AI (GLM provider), paste it into Cursor's OpenAI settings, override the endpoint, and add GLM 5.2 as a custom model.
• Codex Setup: Use OpenRouter, obtain an API key, configure the provider endpoint, and then set up GLM 5.2 as a custom model within Codex.

Benefits of Local Models and Model Chaining

• Key benefit: avoiding token costs associated with cloud-based models.
• Local models allow for continuous task execution without per-token fees, making them cost-effective for startups.
• Model Chaining/Fusion: Combine models to leverage strengths. For example, use a high-end model (like Opus 4.8) for vision tasks (explaining images) and then a more efficient model (like GLM 5.2) for execution based on that analysis.
• This approach offers top-tier output quality at a significantly lower cost.

Cost Considerations and Future-Proofing

• Using GLM 5.2 via OpenRouter can be ~5x cheaper than Opus 4.8 for similar output quality (e.g., 44 cents vs. $2.38 for a large task).
• The cost of AI subsidies on tokens may decrease over time; investing in local hardware now could be cost-effective for future models.
• Companies are re-evaluating token spending due to increasing costs, leading to interest in local solutions and better model governance.

Practical Applications and Recommendations

• GLM 5.2 excels at refining designs and executing instructions, even with limitations like lack of native image understanding.
• For those starting out, using model-agnostic harnesses like OpenRouter and Cursor is highly recommended to experiment with local models without immediate hardware investment.
• Prioritize token minimization and output maximization rather than just raw token usage.

App Idea Generation

• Solve your own problems and build something you're passionate about.
• Ideas should be simple, solve one clear problem, and target a defined audience.
• Leverage AI for unique, valuable, and engaging features.
• Scroll through social media (Instagram, TikTok) to identify audience problems and potential influencer promotion angles.

Product Development & AI

• Use AI tools (like ROR) to build apps without coding knowledge.
• Focus on a core "gotcha" feature that is instantly understandable and provides high value.
• Prioritize simple and intuitive design (your mom test).
• Dedicate 14 days to core functionality and branding, then focus on onboarding and revenue integration.

Distribution & Marketing

• Build a professional Instagram page with demos and a clear Call to Action (CTA).
• Use a VA to handle influencer outreach and initial contact.
• Tailor your "For You Page" (FYP) to your ideal customer profile for efficient lead generation.
• Influencer marketing is key; aim for partnerships, offer equity, and negotiate rates effectively by demonstrating potential.
• Paid ads can be effective for scaling, testing creatives, and replicating successful ad strategies from competitors.

Key Metrics & Optimization

• Track downloads, conversion rates, and Average Revenue Per User (ARPU).
• Analyze retention rates to identify areas for improvement and feature additions.
• Focus on building a useful product; adding features like calorie tracking improved retention for Wrestle AI.

Mindset & Opportunity

• This is a prime time for app building, similar to the e-commerce boom.
• Many people are unaware of AI's capability to build software, creating a window of opportunity.
• Develop strong sales and social skills to effectively close influencer deals and build relationships.

Fable 5 Ban and Cloud Model Vulnerability

• Claude Fable 5 was banned due to a US government letter, highlighting the fragility of relying on cloud-based AI models.
• Businesses built on rented access to AI models are vulnerable to sudden shutdowns by governments or companies.
• The event underscores the need for users to own a part of their AI infrastructure for resilience, similar to having a backup generator.

The Rise of Local AI Models

• Local AI models run entirely on a user's computer, requiring no internet, API keys, or per-token costs.
• They offer enhanced privacy as data never leaves the machine, crucial for industries with strict data regulations (healthcare, legal, finance).
• Local models provide zero marginal cost after initial hardware investment, allowing for unlimited use.
• These models are censorship-resistant and function independently of external services or internet connectivity.
• While historically less powerful, modern local models on gaming GPUs or capable Macs are now sufficient for approximately 80% of common AI tasks.

Getting Started with Local Models

• Runtime: Download an AI runtime like Olama (command-line) or LM Studio (user-friendly interface).
• Hardware Matching: Match model size (billions of parameters) to your hardware. 4B models run on most devices, 12B are suitable for 16GB RAM machines, and 27-35B+ require higher-end setups.
• Key Models:
  • Qwen 3/3.6: Good all-around choice, strong in coding and multilingual tasks.
  • DeepSeek: Excels at complex thinking and coding, but may have a 10-30 second response delay.
  • Gemma: Google's efficient model, with versions fitting on 16GB RAM or phones.
  • Llama: Widely supported by a large community, runs almost anywhere.
• Quantization: A technique to shrink models (e.g., Q4, Q5) for lower hardware requirements with minimal quality loss, making powerful models accessible on consumer hardware.
• Agents: Connect local models to agents like Hermes for enhanced capabilities, creating offline, persistent AI assistants.

Advanced Local Model Usage & Startup Ideas

• Context Window: A primary constraint for local models; larger context windows consume more RAM. Keep sessions focused.
• Tool Integration: Equip local models with tools (web search, file access, code execution) to significantly boost their capabilities.
• Startup Opportunities:
  • On-device AI for regulated industries (healthcare, legal, finance).
  • Local, private versions of existing cloud AI tools.
  • Air-gapped agents for highly sensitive operations.
  • Offline AI for areas with no internet access (ships, rural clinics).
  • "Resilience as a Service" – providing fallback AI solutions for businesses concerned about provider outages.
• The core lesson is to build on a durable foundation and own a part of your AI stack, using local models as insurance against disruptions.

Video Editing with Fable 5

• Fable 5 can be used for professional video editing, including stitching best shots, removing filler words (like "ums"), and color grading.
• A prompt example shows Fable 5 orchestrating transcription (11 Labs), shot selection, JSON file creation, and video assembly (FFmpeg).
• Fable 5 can also handle static design frames, rebuilding PNGs as code, and color grading adjustments.

AI Content Engine

• To use Fable 5 effectively for content creation, provide inputs like origin story, known expertise, offer details, ICP, frameworks, content pillars, funnel structures, and tone.
• Establish a weekly loop for research, niche scanning, topic identification, performance analysis, and hypothesis testing.
• Fable 5 can visually inspect data using Co-work browser and operate autonomously for extended periods.

Efficient AI Usage & Cost Optimization

• The key insight is "low effort is the alpha," meaning using Fable 5 strategically for high-impact tasks rather than routine ones.
• Tools like Factory.ai's Droid can orchestrate tasks, utilizing Fable 5 for complex jobs and other models (like Opus) for routine tasks to manage costs.
• Even Fable 5 "Low" can outperform other models at "High" settings for certain tasks.

Copywriting & Landing Page Creation

• Fable 5 can create compelling website copy through a "landing page tournament" process.
• This involves generating multiple copy variations, having them judged by diverse personas (CFO, distracted founder, competitor, ideal customer, copywriter), scoring each, and merging the best elements.
• The AI can generate realistic judge feedback, revealing critical insights for optimizing conversion rates.

Startup Planning & Spec Development

• Fable 5 excels at in-depth interviews to refine startup ideas and achieve product-market fit (PMF).
• It can push back on vague answers and ask probing questions (e.g., "What habit have you personally failed to keep?"), mimicking expert mentors.
• This process leads to a comprehensive spec document, failure analysis, and a V1 build with a higher likelihood of success.

Advanced Fable 5 Use Cases

• Kill Your Company: Use Fable 5 to identify and rank threats to your business, simulating a competitor's strategy.
• Personal Operating Manual: Analyze years of your notes, decision logs, and post-mortems to identify patterns in your decision-making and create a COO-style manual.
• Identify Missing Elements: Ask Fable 5 to identify what's absent from your business by analyzing successful competitors and market gaps.
• Negotiation Pro: Simulate negotiations by having Fable 5 embody the counterparty, understanding their incentives and pressure points, and revealing what you give away.
• Contract Review: Fable 5 can analyze legal documents, leases, and contracts, identifying hidden costs, missed opportunities, and necessary protective clauses, flagging areas needing a human lawyer.
• Self-Automation: Instruct Fable 5 to study its own past interactions, build tools for repeated tasks, and identify areas you should delegate.

Startup Ideas Enabled by Fable 5

• Synthetic Focus Group Firm: Use Fable 5 to run "copy tournaments" based on customer reviews, providing brands with validated ad concepts and customer insights.
• Custom Software in 48 Hours: Build niche internal tools for businesses (e.g., med spas) rapidly and affordably by using Fable 5 to conduct sales interviews, write specs, and generate code.
• Contract Refund Firm: Analyze vendor contracts and invoices to find savings (auto-renewals, unused services) and charge a percentage of the recovered funds. Fable 5's ability to process hundreds of PDFs is key.

What is an Agentic Loop?

• Human in the Loop: Traditional AI interaction where a human prompts an agent, reviews the result, and iterates. The human is the primary director.
• Agentic Loop: A system where an AI agent generates a result, feeds it back to itself as input, and continues to work without direct human intervention after the initial prompt.

The Hype and Dangers of Agentic Loops

• Assumption Problem: Agents in a loop make assumptions, similar to a solo developer building a product without client feedback, which may not align with the intended vision.
• Token Burn: These loops can be extremely expensive, consuming vast amounts of tokens, making them impractical for most users, especially those on limited budgets.
• Misaligned Expectations: It's nearly impossible for a human to pre-define every detail for an agent in a complex project, leading to errors and wasted resources.
• "Slot Machine" Analogy: Agentic loops can feel like a gamble, producing unpredictable and often incorrect results due to the lack of human oversight.

When Agentic Loops Can Work

• Experimentation and Prototypes: Useful for quickly building prototypes or small tools where precise details aren't critical.
• Confined, Goal-Oriented Tasks: Effective in scenarios with a clear, binary output and a well-defined feedback loop.
• Code Review Example: The speaker uses a loop for code review where an AI agent (Gravile) scores code, and another agent (Cursor) refines it based on the score until it reaches a satisfactory level (e.g., 4/5 or 5/5). This is constrained by line limits and specific objectives.
• SEO Generation: Can be useful for generating large volumes of similar content, like SEO pages, where creativity is not a primary factor.

The Future of Agentic Loops

• Not Yet Practical for App Building: While the concept may be the future, it's currently not a reliable or cost-effective method for building complex applications or startups.
• Human Oversight is Key: For creative tasks and building meaningful products, the "human in the loop" remains the most effective and efficient approach.
• Potential for Later Development: The technology is advancing, and agentic loops might become more viable in the future, but not in their current state for most use cases.

Becoming AI Native

• An AI native org is one where people manage agents, agents can read/write to the company, and the company gets smarter over time.
• This system unlocks speed (creating outputs in minutes) and signal (real-time market feedback).
• Key components are people (strategy, judgment, trust), agents (using tools in a loop), and context (readable data for agents).

The Role of People & Agents

• AI handles execution ("the middle"), freeing people to focus on strategy and review.
• Everyone is now a manager of agents, responsible for setting them up for success.
• Autonomous agents need a clear goal, skills, tools, and context.
• Evals are crucial for visibility into agent output and quality.

Context & Skill Chains

• The context layer ("brain") provides agents with "20/20 vision" of the company's data.
• Information is captured from various sources, curated, and stored for agent access.
• Skill chains are sequences of skills that allow agents to perform complex tasks, like generating a personalized proposal in minutes.
• This system combats AI "hallucinations" by grounding outputs in specific data.

Practical Applications & Startup Ideas

• A demo showed a skill chain creating a detailed, personalized proposal microsite in minutes.
• Another demo showcased building a functional prototype for Spotify's "daily blitz" feature in under 10 minutes, including user testing and feedback synthesis.
• This rapid prototyping and feedback loop is game-changing for product development.
• Startup opportunities lie in applying the AI native framework to specific industries, functions, and company sizes, especially focusing on niche, high-frequency workflows.

Hermes Desktop App Launch & Advantages

• Hermes Desktop app is presented as a superior alternative to using AI agents via Telegram, offering enhanced productivity and a better user experience.
• It introduces the concept of 'sessions' to manage conversations and context, preventing context pollution and reducing costs, unlike the single-thread approach common in Telegram.

Profile Management & Cost Savings

• Hermes Desktop allows for the organization of multiple 'profiles,' each representing a distinct AI agent with its own personality, skills, and memories.
• Users can efficiently switch between profiles tailored for specific tasks (e.g., GPTM for coding, Opus for high-level strategy, Quen for local research), optimizing for cost and performance.
• This profile system is preferred over role-based agents (like a dedicated product manager AI) for cost efficiency and direct task execution.

Key Features for Productivity

• Artifacts: A centralized location to manage all saved links, images, media, and files, acting as a 'second brain' for organized information retrieval.
• Skills & Tools: An interface to manage over 150 pre-installed skills, allowing users to disable unused ones to save costs and see newly generated skills based on usage.
• Messaging Services: Easy setup for integrating messaging platforms like Telegram without needing to use the command-line interface (CLI).
• Cron Jobs: A dedicated section to view, manage, and create scheduled tasks (cron jobs) with high confidence in their execution, eliminating the unreliability often found with CLI-based scheduling.

Advanced Techniques & Use Cases

• Reverse Prompting: A technique for creating highly effective prompts by asking the AI itself for the best prompt to achieve a desired outcome, especially useful for complex tasks like setting up cron jobs.
• Sub-agents vs. Profiles: Profiles are for agents with different skill sets, while sub-agents are copies of the main agent used for performing multiple instances of the same task simultaneously.
• Making Money with Hermes: A core use case involves using cron jobs to automate scanning for challenges and opportunities on platforms like Reddit and X. The agent identifies problems, explains why the user is positioned to solve them, and can even auto-generate prototype solutions.

Hardware Recommendations

• For running local models efficiently, the DGX Spark is recommended as a plug-and-play device with significant unified memory.
• The Mac Studio is also a good option, though often sold out and with less memory in standard configurations.
• Investing in local hardware is advised due to rising memory and hardware costs, viewing it as an investment for learning and generating ROI.

Introduction to Codex Sites

• Codex Sites offer autonomous product building, unlike static sites that require manual updates.
• They are best for those within the Codex ecosystem interested in autonomous app development.
• Current limitations include lack of public domain publishing, databases, payments, email sending, analytics, and secret vaults.

Building with Codex Sites: A Startup OS Example

• To start, invoke "sites" as a plugin in Codex.
• Key prompts: "build a startup OS," "use realistic sample data," and crucially, "save for review, do not deploy."
• Plugins like Figma, Canva, HeyGen, and Game Studio can enhance site interactivity.

Adding Memory and Safe Actions

• Prompt for persistence storage to enable ideas to save between visits.
• Codex Sites can use Cloudflare D1 for storage, defining data models (e.g., "ideas" with fields like title, column, score).
• "Safe actions" are approved buttons or commands that automate app modifications, preventing arbitrary changes.
• Example safe actions: add idea, update idea.

Creating Skills and Gatekeeping

• Create a Codex skill (e.g., "startup ideas admin") to provide instructions and example commands for using the app.
• Skills act as reusable instruction manuals for Codex.
• "Save gate" or checkpoint your work using prompts like "save this as v1 review, do not deploy" to create reviewable versions without deploying.

Autonomous Operation and Publishing

• "Proving the loop" involves using a skill in a new chat to add data to the app, demonstrating autonomous updates.
• The app can be published to a temporary URL, showcasing its functionality.
• The core value lies in creating products that Codex can continuously operate and update autonomously.

Key Concepts for Codex Sites

• Memory: Essential for the app to save data and function beyond a demo.
• Safe Actions: Approved methods for automating app edits, enabling autonomous updates.
• Skills: Reusable instruction manuals that Codex uses to operate the app.

The Agent Web vs. The Human Web

• The internet is shifting from human users to AI agents as customers.
• The human web focused on persuasion for human attention.
• The agent web requires machine usability, structured capability, permission, and trust.

The Agent Buying Journey

• Agents will discover, evaluate, invoke tools, pay, and renew services.
• This journey includes finding, evaluating (docs, pricing, APIs), trusting (policies, limits), transacting (paying, booking), using tools, and recommending other tools.

Infrastructure Needs for Agents

• Agents require identity, tools, an inbox, memory, a wallet (with spend caps/approvals), and receipts.
• This is analogous to an employee's increasing trust and access.

Concrete Examples and Opportunities

• Agent Mail: Inboxes for AI agents (e.g., email API for agents).
• Fintech: Agents with wallets for purchasing software, with spend caps and approvals.
• Support: Agents filing tickets, requesting refunds, and following up.
• Procurement: CFO agents comparing vendors and negotiating terms.
• Local Business: Travel agents booking and managing reservations.

Making Your Business Agent-Readable

• Shift from SEO to AEO (Agent Enablement Optimization).
• Websites need structured docs, schemas, policies, examples, and endpoints.
• Provide MCP tools, SDKs, OAuth, checkout, sandbox, and receipts.
• Optimize for agent actions rather than human persuasion.

Startup Ideas for the Agent Era

• Agent SEO agency, agent identity and permissions tech, agent receipts/audit trails, agent-ready doc generators, agent inbox security, agent-readable pricing pages, MCP servers for franchises, agent support desks, and agent sandboxes for testing SAS.

Biggest Creator Opportunities

• Live Shows & Unscripted Content: The future of creators involves live, unscripted content, similar to gaming streamers on Twitch. This human element is a counterpoint to AI-generated content.
• Monetization: High affinity audiences can monetize through events, merchandise, and platforms like Patreon, even with a smaller viewership.

Action Apps (AI Agents First)

• Concept: Apps that perform tasks on behalf of the user, rather than requiring user interaction. Think AI agents managing inboxes or calendars.
• Opportunity: Companies resistant to transitioning to an "agent-first" model create an opportunity for new players to build these applications.
• User Experience: While full automation is the goal, some user interaction might be needed to maintain perceived value and engagement.
• Development: Leverage existing SDKs (like Cloud Agent SDK) to build "agent wrappers" on top of AI models.

Solving Loneliness & Building Community

• Problem: Increasing loneliness across generations presents a significant societal and business opportunity.
• Solutions: This includes "third spaces," community apps, and in-person gatherings (e.g., retreats, hobby clubs).
• Niche Communities: Highly specific online communities (e.g., "Dads of Marathon" Discord) demonstrate a strong thirst for connection.
• Business Models: Membership models and event organizing businesses (e.g., 222, Fabric) are viable.

Elder Tech

• Market: A large, underserved demographic (65+) with significant needs in areas like hearing, mobility, social connection, memory, and vision.
• AI Integration: AI can greatly enhance solutions for these challenges.
• Targeting: Older adults want products that improve their lives, not necessarily ones that explicitly market to their age.
• Untapped Potential: Many older adults have disposable income and are often overlooked by product developers.

Hobbies & Adult Learning for Joy

• Demand: Adults, particularly those whose children have left home or are seeking change, are looking for activities that bring joy, not just professional development.
• Examples: Painting, woodworking, pottery, and other creative pursuits are in demand.
• Low Competition: These areas often have less competition as they're perceived as less lucrative than tech or business ventures.
• Monetization: Premium retreats and workshops for specific niches (like entrepreneurs) can be highly profitable.

AI Employees

• Concept: Businesses can offer "digital workers" or AI agents to perform specific tasks for other companies, particularly in white-collar roles.
• Niche Focus: Verticalization (e.g., AI for accountants in Germany) and focusing on specific job titles (e.g., junior YouTube editor) is key.
• Value Proposition: Clearly demonstrate how AI employees can reduce costs and workload, especially for menial or repetitive tasks.
• Branding: Positioning AI employees as assistants for junior-level tasks, rather than replacements for senior creative roles, can build trust.

Personalized Nutrition

• Data-Driven: Utilizing blood work, gut biome data, and DNA tests to provide tailored dietary advice.
• Verticalization: Focusing on specific health conditions (e.g., GERD, migraines) is more effective than generic health advice.
• Holistic Approach: Businesses could offer a comprehensive solution, from testing to personalized meal plans or chef recommendations.
• Market Indicators: Over-the-counter medication aisles (e.g., for GERD) highlight prevalent problems.

Pet Health Tech

• Massive Market: The pet industry is large, with owners increasingly treating pets as family members.
• Smart Monitoring: Integrating AI and hardware for health monitoring (heart rate, sleep) offers significant opportunity.
• Innovation: Apply human health trends and technologies (e.g., supplements, AI analysis) to pet care.

AI-Native Media Companies

• Strategy: Utilize AI to build large audiences on social media platforms, with transparency about AI usage.
• Quality is Key: "AI slop" is unlikely to succeed long-term; high-quality, niche content is essential.
• Audience Monetization: Once an audience is built, leverage it to sell products, apps, or services.
• Impactful Application: Even AI-generated content can have a positive impact if it helps build an audience for valuable expert content or solutions.