Key Moments
Inside YC's AI Playbook
Y CombinatorWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
Companies are building "superintelligence" by making AI the company's operating system, not just a feature, by giving it unrestricted database access and creating self-improving skill loops.
Key Insights
YC started building internal AI agent infrastructure about a year ago, initially focusing on giving the finance team control over their own workflows using natural language prompts instead of code.
Giving agents unrestricted read-only SQL query access to YC's single, centralized PostgreSQL database containing all company data was a major unlock, enabling arbitrary business questions to be answered.
The number of complex questions asked dramatically increased due to the ease of access; what previously took hours of SQL writing now takes minutes or less.
YC's internal tool registry has grown to over 350 tools, enabling teams to add agent-specific capabilities ranging from managing office hours to booking journal entries.
Self-improving loops, like an agent that reviews past conversations to identify areas for improvement, are enabling a 'shared organizational brain' that gets smarter overnight.
The future of AI development points towards agent-wrapped deterministic tools rather than deterministic software wrapping AI, with chat as a powerful, intuitive interface.
Integrating AI as the organizational operating system
Building "superintelligence" within a company is not about adding AI as a supplementary feature, but about making it the fundamental operating system that the entire organization runs on. This involves recording "all the artifacts" of AI interactions, essentially creating a "shared organizational brain." Pete Koomen, who led the development of YC's internal agent infrastructure, emphasizes that this approach allows every employee to leverage the collective skill and instinct of their colleagues, transforming a pre-AI organization into an AI-native one. The journey began with a focus on empowering non-technical teams, like finance, to manage their own workflows through natural language, bypassing traditional software development loops.
Unlocking AI power with unrestricted database access
A pivotal moment in YC's AI development was granting agents unrestricted read-only access to the company's singular PostgreSQL database. This database houses all critical YC information, including details on funded companies, founders, financial transactions, and internal notes. Previously, answering complex business questions required significant effort from software engineers and data teams. However, with direct SQL query capabilities for agents, arbitrary questions like "show me all investors who invested in a space-related company in the last four batches" could be answered almost instantaneously. This democratized data access dramatically increased the frequency and complexity of questions asked, resolving the bottleneck of inter-team dependency. This mirrors early advancements like Google's "big table" concept, which involved denormalizing data into a single, accessible format for efficient processing.
The evolution of YC's internal tool registry
Starting with a simple agent loop and a few tools, YC has grown its internal tool registry to over 350 specialized tools. This registry is crucial for adapting general AI agents to specific organizational needs. Teams can add tools for tasks like managing office hours, booking journal entries, or assisting with event management. These tools transform agents into powerful, work-specific assistants. Importantly, this registry creates a shared resource. While internal agents utilize these tools, they can also be exposed to individual developer tools like Cloud Code running on personal machines, bridging the gap between organizational and individual AI capabilities. This comprehensive registry ensures that the most relevant and specific functions are available to the AI systems.
Self-improving skill loops and the 'organizational brain'
A key innovation is the development of self-improving loops, where agents learn and evolve autonomously. One notable example is a general agent that reads through all internal employee conversations nightly to identify areas for improvement or contexts that would have led to more efficient outcomes. This process is akin to Karpathy's "auto-research" or the "dream cycle" found in tools like OpenClaw and Gbrain. These loops can even write back insights into internal databases or CRMs. This continuous learning mechanism is crucial for building what is described as a "shared organizational brain," where the collective knowledge and experience of the company are constantly being refined and integrated.
The "two-sentence description" skill as a microcosm of AI-native growth
The development of a "two-sentence description" skill for YC founders serves as a powerful illustration of how AI can drive organizational intelligence. This skill, initially crafted by a partner, condenses complex company information into a concise, understandable pitch. Through analyzing meeting transcripts where partners provided feedback on founder pitches, the agent learned and improved the skill significantly, eventually surpassing human expertise in its specific domain. This process highlights how an organization can capture tacit knowledge, refine it through AI, and create a system that is better than any individual. It underscores the concept that superintelligence within a company is achieved by systematically applying this improvement process to thousands of similar atomic tasks.
The shift from "horseless carriages" to AI-native software design
Pete Koomen's "Horseless Carriages" essay critique argued against simply adding AI features into existing software (like an AI email writer). Instead, the true potential of AI lies in shifting control from developers to users and fundamentally rethinking software architecture. This means agent-wrapped deterministic tools, rather than deterministic software wrapping AI. Chat interfaces are emerging as a powerful and intuitive way to interact with these agents, as they closely mirror human language and thought processes. This approach allows for "just-in-time" software, where agents can generate tailored interfaces or components precisely when needed, moving away from rigid, pre-built applications.
The "Apple 1 moment" for personal AI and decentralization
The development of AI presents a critical choice: a centralized future controlled by a few large entities, or a decentralized one empowering individuals. The centralized model mirrors the era of mainframes, where access to computing power was restricted. The vision for a decentralized future, akin to the "Apple 1 moment" with personal computers, involves primitives that allow individuals to run their own software, customize prompts, and maintain private repositories. Tools like Gbrain and OpenClaw are seen as steps towards this "personal AI" revolution, enabling users to control their AI as an extension of themselves, rather than being dictated by corporate interests. This contrasts with the limited access often provided by mainstream platforms like ChatGPT.
AI as an enabler of empowerment, not replacement
Contrary to fears of AI replacing human jobs, the experience at YC and in the broader tech industry suggests AI is a powerful tool for individual empowerment. Just as personal computers and the internet democratized publishing, AI is poised to eliminate tedious, "drudgery-style" work, allowing individuals to focus on more impactful tasks. However, achieving this requires deliberate organizational choices, such as fostering trust, openness, and egalitarian access to AI tools for all employees, not just leadership. This requires a radical shift from traditional command-and-control structures. The decisions made now about how AI is integrated will shape its future impact on society, emphasizing the importance of building AI as an extension of human capability.
Mentioned in This Episode
●Software & Apps
●Companies
●Books
●Concepts
●People Referenced
Common Questions
Companies can build superintelligence by not just using AI as a copilot but as the fundamental building layer for everything. This involves recording all artifacts, using AI to encode workflows in natural language, developing internal tool registries, and creating autonomous, self-improving loops that allow AI to enhance its own capabilities and knowledge.
Topics
Mentioned in this video
General Partner at YC, creator of Optimizely, and leader of YC's internal AI agent infrastructure development.
Mentioned for his vision of block becoming a mini-AGI focused on payment processing.
Co-founder of Apple, mentioned for his role in the personal computer revolution.
Co-founder of Apple, mentioned for his role in the personal computer revolution.
A company co-founded by Pete Kooman, known for its early work in A/B testing for apps and websites.
A powerful agentic tool that allows extensive customization and access to systems, contrasting with more restricted software.
An agentic tool mentioned alongside OpenClaw and other harnesses, designed for powerful interactions.
A coding agent harness mentioned as an example of single-player agent tools.
The company led by Jack Dorsey, aiming to become a mini-AGI for payment processing.
Mentioned in the context of the personal computer revolution and the choice between centralized and decentralized AI.
A tech giant whose interests might dominate AI development, contrasted with the vision of individual control and programming.
A tech giant whose interests might dominate AI development, contrasted with the vision of individual control and programming.
A major AI research lab whose interests might compete with individual control over AI development.
A major AI research lab whose interests might compete with individual control over AI development.
The widely used AI model, presented as an example of more centralized AI with limited customization compared to open-source alternatives.
An early agentic coding tool mentioned as being well-established around the time YC began developing its own infrastructure.
A tool that was newly introduced or gaining traction, contrasted with older software development methods at YC.
The database system where YC stores all its critical organizational data, enabling powerful agentic queries.
Business intelligence tools that were previously used for data analysis, contrasted with the efficiency gains from AI agents querying the database directly.
An open-source system similar to Gary's List 2.0, designed for agentic retrieval and knowledge management.
Mentioned as a system that can wrap agents, though CLI and direct interaction seems more effective for some agents.
Command Line Interface, noted as an efficient way for agents to interact with systems.
A coding agent harness mentioned as an example of single-player agent tools.
A communication platform used internally at YC to broadcast agent conversations and enable learning.
Used as an example of a product that added AI features in a way that locked down prompt context from the user (safetyism).
An AI model mentioned as being slightly more open than ChatGPT but still limited compared to tools like OpenClaw.
An AI search engine that offers a better version of centralized AI but is still limited compared to open-source agents.
An open-source agent framework that allows for significant user control and customization.
A programming language used in modern agentic frameworks, offering more dynamic capabilities than traditional frameworks like Rails.
More from Y Combinator
View all 592 summaries
68 minInference, Diffusion, World Models, and More | YC Paper Club
51 minWhy Good Companies Go Bad (And How to Stop It)
29 minWhy Zepto's Aadit Palicha Turned Down Stanford to Deliver Groceries
22 minPaul Graham, Founder of Y Combinator, Live from Stockholm
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free