Quick Read
Summary
Takeaways
- ❖YC transitioned from AI as a copilot to AI as the core building layer for all internal operations.
- ❖A single PostgreSQL database housing all YC's critical data (companies, founders, finance, CRM notes) was foundational for agent effectiveness.
- ❖YC developed an internal tool registry, growing to over 350 tools, allowing agents to perform YC-specific tasks like managing office hours or booking journal entries.
- ❖Non-technical teams, like finance, gained control over their software by encoding workflows with English prompts, bypassing traditional engineering loops.
- ❖The ability for agents to run read-only SQL queries against the unified database dramatically increased the number and complexity of business questions YC could answer.
- ❖AI agents now self-improve by analyzing past employee conversations, leading to skills (like writing two-sentence company descriptions) that surpass individual human expertise.
- ❖Transparency, with agent conversations broadcast internally, fostered learning and acted as a social control for security in a high-trust environment.
- ❖Building an AI-native organization requires an egalitarian, trust-by-default culture and a willingness to invest significantly in token usage.
- ❖The future of software is 'just-in-time,' with agents dynamically generating purpose-built applications based on user needs and prompts.
- ❖AI can be either centralizing (locked-down, developer-controlled) or decentralizing (user-controlled, empowering); YC advocates for the latter.
Insights
1Unified Data Context is Foundational for AI Superintelligence
YC's ability to create powerful internal AI agents stems from running all its operations on proprietary software, consolidating every piece of critical data (companies, founders, finances, CRM notes) into a single PostgreSQL database. This unified context allows agents to answer complex, arbitrary questions about the business that would otherwise require hours of manual SQL or cross-team coordination.
YC runs on its own software, all of which sits on one PostgreSQL database that has everything important to YC's world in it. This enables agents to answer questions like, 'Show me all the investors who invested in a space-related company in the last four batches.'
2Shared Tool Registries Empower Non-Technical Users and Drive Adoption
Beyond a unified database, YC developed an internal tool registry, initially with 20 tools, now exceeding 350. These YC-specific tools allow agents to perform functions like booking journal entries or managing office hours. This system empowers non-technical teams, such as finance, to encode and automate their own complex workflows using natural language prompts, removing engineers from the loop and making software creation accessible.
The tool registry is where most of the YC-specific stuff lives. It turns these agents into something useful at work. Teams have added more and more tools, now over 350, allowing finance to book journal entries or partners to manage office hours.
3Self-Improving Skills Lead to Organizational Superintelligence
YC's agent system incorporates autonomous, self-improving loops. A general agent reads through employee conversations, identifies areas for improvement, and uses this context to refine existing skills. This process has led to AI skills, like the 'two-sentence description' generator, becoming more effective than individual human partners, creating a collective, continuously improving organizational intelligence.
We have this general agent that every night will go and read through all of the agent conversations that employees have had and look for things it could have done better... improve the two-sentence description skill and they got noticeably better after that like this thing is now better than I am I would I would argue at writing those.
4Transparency and Trust are Critical for AI-Native Culture
YC's internal policy of making agent conversations globally viewable by all full-time employees, broadcast to a Slack channel, fostered rapid learning and adoption. This 'trust-by-default' and egalitarian approach, while initially debated for security, created a social control mechanism and allowed employees to learn from each other's creative AI usage, accelerating the organization's AI transformation.
By default the agent conversation is actually globally viewable by any full-time employee at YC... People learned how to use it from watching how other people used it. We used that transparency to solve several problems at the same time.
Bottom Line
AI's trajectory presents a choice between centralization and decentralization, mirroring the mainframe vs. personal computer era. A centralized future, where a few 'kings' control AI and user prompts are locked down, is a risk. The alternative is a 'personal AI moment' akin to the Homebrew Computer Club, where individuals control their own AI, prompts, and models.
Organizations and individuals must actively choose to build and support decentralized AI systems that empower users with control over their data and tools. This requires fostering open-source initiatives, customizable agents, and the ability to run personal AI on private infrastructure, rather than relying solely on large, proprietary models with locked-down interfaces.
Develop open-source, customizable agent frameworks and personal AI platforms that prioritize user control, data privacy, and the ability to integrate with diverse data sources. Build tools that enable individuals and small teams to create and manage their own AI 'brains' and skill registries, fostering a new wave of innovation outside of corporate giants.
The 'chat as the interface' model, initially doubted, is proving to be the most effective for AI applications because it's the closest to human language and thought expression, allowing for multimodal inputs (text, voice, images, files) and dynamic, 'just-in-time' software generation.
Developers should lean into chat-based interfaces for AI products, focusing on robust multimodal input capabilities and the agent's ability to dynamically interpret and execute complex commands. Resist the urge to over-engineer rigid UIs, as the power lies in the flexibility of natural language interaction and the agent's capacity to generate bespoke solutions.
Create frameworks and platforms that simplify the development of multimodal, chat-native AI applications capable of generating 'just-in-time' software components. Focus on improving the agent's ability to understand context, user intent, and dynamically create UI elements or execute complex operations through natural language commands, reducing the need for traditional, static front-end development.
Opportunities
AI-Powered 'Just-In-Time' Software Generation Platform
A platform that enables non-technical users to describe desired software functionalities in natural language, and AI agents dynamically generate and deploy small, purpose-built applications (e.g., single-page JavaScript apps, custom reports) on demand. This eliminates the need for extensive coding and provides immediate, tailored solutions.
Decentralized Personal AI Operating System
Develop an open-source operating system or framework that allows individuals to run and control their own AI agents, choose open-weight models, manage private data repositories, and customize prompts and skills. This would be the 'Apple 1 moment' for personal AI, empowering users outside of corporate ecosystems.
Organizational AI Onboarding & Apprenticeship System
A system that leverages an organization's collective AI 'brain' (unified data, skill registry, recorded conversations) to rapidly onboard new employees. The AI acts as a perpetual mentor, answering 'dumb questions' new hires are embarrassed to ask, simulating expert interactions, and teaching best practices from top performers, significantly reducing ramp-up time.
Key Concepts
Shared Organizational Brain
The concept of consolidating all an organization's internal data and knowledge into a single, accessible system (like YC's unified PostgreSQL database and tool registry) that AI agents can leverage to answer arbitrary questions and automate tasks, effectively creating a collective intelligence accessible to all employees.
Jevons Paradox (AI Context)
When the efficiency of a resource (like answering complex business questions) increases dramatically due to AI, the demand for that resource (the number and complexity of questions asked) also increases, leading to greater overall consumption and insight generation, rather than just doing the same amount with less effort.
DRY and MECE Resolver
A principle for organizing AI skills and tools in a registry to be 'Don't Repeat Yourself' (DRY) and 'Mutually Exclusive, Collectively Exhaustive' (MECE). This ensures that skills are unique, cover all necessary domains, and are optimally structured for AI models to understand and utilize efficiently, preventing redundancy and improving agent performance.
Just-In-Time Software
The idea that AI agents can dynamically generate and build purpose-built software or interfaces on the fly, tailored precisely to a user's immediate need or query, rather than relying on pre-built, rigid applications. This allows for extreme flexibility and responsiveness, making the chat interface a powerful conduit for complex interactions.
Lessons
- Consolidate your organization's critical data into a single, unified database to create a comprehensive context layer for AI agents.
- Build an internal tool registry with YC-specific tools, allowing teams to encode and automate their workflows using natural language prompts, reducing reliance on engineering.
- Implement a 'trust-by-default' culture with transparent AI agent conversations (e.g., broadcasting to internal channels) to foster learning, accelerate adoption, and create a social control for security.
- Invest in self-improving AI loops where agents analyze past interactions to refine and enhance existing skills, leading to continuously improving organizational intelligence.
- Prioritize empowering users with control over AI tools and prompts, moving beyond 'AI as a copilot' to 'AI as the building layer' for all operations.
Building an AI-Native Organization: YC's Playbook
**Establish a Unified Data Context:** Consolidate all critical organizational data (CRM, finance, HR, project notes) into a single, accessible database (e.g., PostgreSQL). This serves as the 'organizational brain' for AI agents.
**Develop a Shared Tool Registry:** Create an internal registry of specific tools that allow AI agents to interact with your organization's systems and perform discrete tasks. Encourage every team to contribute and expand this registry.
**Empower Users with Agentic Workflows:** Enable non-technical employees to define and automate their workflows using natural language prompts through these agents, bypassing traditional software development cycles.
**Foster Transparency and Trust:** Default to making AI agent conversations and tool usage transparent and viewable across the organization. This accelerates learning, adoption, and establishes a high-trust environment.
**Implement Self-Improving AI Loops:** Design agents that can analyze past interactions, identify inefficiencies, and automatically refine their own skills and prompts, leading to continuous improvement and 'superintelligence'.
Notable Moments
The initial impetus for YC's internal AI project came from inefficiency in building software for the finance team, where engineers had to deeply understand complex financial workflows to build deterministic tools.
This highlights a common pain point in organizations that AI agents, by allowing non-technical users to encode workflows with natural language, can directly address and resolve.
The 'magic moment' when YC's agents gained the ability to run read-only SQL queries against their single, comprehensive PostgreSQL database, allowing them to answer arbitrary business questions.
This demonstrated the immense power of a unified data context combined with agentic capabilities, dramatically increasing the scope and frequency of data-driven insights.
The realization that AI-generated 'two-sentence descriptions' for companies became better than those written by experienced YC partners after the agent learned from meeting transcripts and partner feedback.
This is a concrete example of how self-improving AI skills can lead to organizational superintelligence, surpassing individual human expertise through collective learning.
The decision to make all internal AI agent conversations globally viewable by YC employees, broadcast to a Slack channel.
This radical transparency, initially a point of debate, proved crucial for accelerating learning, adoption, and establishing a social control mechanism for security in a high-trust environment.
Quotes
"Part of the key thing is not to just use AI as a copilot. This is the the thing where you use it as the building layer for everything. And you need to start recording all the artifacts."
"It's like a shared organizational brain. It's like the closest thing to us being able to like connect our brains."
"It just turns out when all of that context is in one place with a little bit of additional information about how the schema is laid out, an agent can go and ask any or answer arbitrary questions about about our business."
"This is how super intelligence happens inside organizations. I mean this two sentence pitch thing sounds like something kind of small but uh embedded in it is actually something very powerful."
"If you frame this as a way for everyone in an organization to get better at what they do using the like collective skill and instinct of of the people they work with. It's incredibly powerful."
"You have to be relatively egalitarian and you also have to be trust by default. And then neither of those things uh actually are most organizations in the world."
"The potential for AI is to shift control of software from the developer to the user."
Q&A
Recent Questions
Related Episodes
Inference, Diffusion, World Models, and More | YC Paper Club
"This YC Paper Club session explores cutting-edge AI research, from accelerating large language model inference and building robust world models for robotics to understanding deep learning generalization and optimizing pre-training under data constraints."
How The Best Companies Defend Against Mediocrity And Rot
"Eric Ries, author of 'The Lean Startup,' reveals how companies can build 'governance fortresses' to protect their core mission from the corrosive forces of shareholder primacy and conventional 'best practices,' ensuring long-term value creation."
Thin Harness, Fat Skills: The New Way To Build Software
"Gary Tan, YC CEO, details how he shipped hundreds of thousands of lines of code in months after a 13-year hiatus, leveraging AI agents and a "thin harness, fat skills" approach to achieve 400x productivity."
How Razorpay Became India’s Largest Payments Company
"Harshil Mathur, co-founder of Razorpay, details the company's path from a YC-backed startup to India's largest payments platform, highlighting critical pivots, the strategic advantage of regulated markets, and the power of customer trust."