How to Build Superintelligence Inside Your Company
In a Nutshell
The core message is that organizations can build superintelligence by treating AI as the foundational infrastructure layer rather than a copilot—recording all artifacts, consolidating data into unified systems, and creating shared agent skills that compound collective knowledge. Key enablers include a tool registry with 350+ functions, transparent agent conversations visible to all employees, and self-improving skills that automatically extract and codify successful workflows. This approach raises the organizational floor by letting new employees instantly access institutional context while requiring high trust, openness, and willingness to spend on tokens to leapfrog legacy competitors.
These notes were generated by AI and may contain inaccuracies.
The key is using AI not just as a copilot, but as the foundational building layer for everything. All artifacts must be recorded to create a shared organizational brain that connects collective skill and instinct across the organization. This approach enables everyone to improve using the combined capabilities of their colleagues.
Pete Kumman, General Partner at YC, created Optimizely, one of the first and best AB testing platforms for apps and websites. He subsequently built all of YC's agent infrastructure including harnesses and internal AI systems. YC has been funding primarily AI companies for the last few years and has iterated through multiple versions of advice for building AI-native companies. Internally, YC has been adopting the same tools and transforming from a pre-AI organization into a super AI-native organization.
The project began about a year ago when Pete and a few engineers started building YC-specific agents and harnesses. This infrastructure has since expanded with contributions from the broader engineering team and partners. The original impetus came from work with the finance team on building tools for workflows like booking journal entries and logging priced rounds.
The traditional loop involved finance teams describing complicated workflows to software engineers, who would then build deterministic purpose-built software, creating an inefficient cycle. At the same time, agentic coding tools like Windsurf, Cursor, and Claude Code were emerging and providing individual superpowers. This created a growing divide between the classical software-building approach at YC and the new agentic methods being used personally.
The solution was to build tools allowing agents to give the finance team direct control over their own software, removing engineers from the workflow loop. Workflows would be encoded in English through prompts rather than Ruby code. This approach began with LLMs for writing SQL queries before agent coding became widespread. Non-technical but intelligent finance team members could successfully use these tools to ask real questions about the business.
A key unlock came from tools providing agents with the ability to run read-only SQL queries against YC's production database and read model files. The database access was initially pushed out despite concerns about security, as earlier tools had been too narrowly scoped. This worked extremely well and demonstrated that reducing security and privacy concerns could unlock powerful capabilities.
YC's major advantage was running entirely on its own software sitting on a single Postgres database containing all important context: companies table, founders table, financial transactions, and internal CRM notes. Most companies outsource functions to third-party SaaS tools, fragmenting their data. YC's unified database enabled agents to answer arbitrary business questions, such as showing all investors who invested in space-related companies in the last four batches.
The database access didn't just make answering questions easier. It dramatically increased both the number of questions asked and the scale and complexity of questions considered worth pursuing. Previously, using BI tools to answer questions like investor activity in specific sectors required several hours of SQL writing, so such questions were rarely asked unless critically important. This exemplifies Gvon's paradox where removing back-and-forth between teams increases the volume of work attempted.
Companies in legacy environments can accelerate by focusing on consolidating context into a data warehouse. This mirrors historical developments like Google's Bigtable, which moved from schema and joins to a single large table optimized for map-reduce. The same principle applies to Karpathy-style knowledge LLM wikis with Gbrain, where data is denormalized into a schema optimized for agent retrieval and understanding. This requires retrieval, RAG, graph RAG, hybrid RRF, and re-ranking capabilities.
Current popular harnesses like Claude Code, Codex, Pi, OpenClaw, and Hermes are designed for single human use on a single machine, making individuals incredibly powerful. The unsolved problem is the multiplayer harness that enables superpowers at team or organizational scale. For legacy organizations, key focus areas include building a common context layer and an internal tool registry.
YC began with an agent loop, simple tool registry, and model router. The tool registry contains all YC-specific functionality that transforms agents into useful workplace tools. Starting with 20 tools including database querying, the registry has grown to over 350 tools as every team adds their own. Examples include office hours management, journal entry booking, and event management tools. These tools are available both to internal agents and to Claude Code running on individual machines.
Skillify is a meta-skill that automatically converts successful agent actions into reusable skills. The resolver pattern appears across systems: agents.md files listing agent capabilities with links to markdown entry points, Claude Code's skill registry functioning as a resolver, and YC's tool registry operating similarly. A meta-skill called check resolvable prevents duplication by ensuring new skills are DRY and MECE (mutually exclusive, collectively exhaustive).
The current period represents simultaneous discovery of useful applied concepts, similar to the early development of Unix with stack and heap concepts. The same primitives keep emerging across different systems: agent loops, tool registries, and skill registries appear consistently whether in Claude Code, internal harnesses, OpenClaw, or Hermes.
Skills function as abstraction layers over tools. YC has developed shared skills accessible through the agent system. The progression mirrors individual development: starting with system prompts, moving to writing skills, then to metaprompting where agents write their own skills. YC has implemented autonomous self-improving loops where a general agent nightly reviews all employee agent conversations to identify improvements and missing context.
A two-sentence description explains what a company does in natural language that anyone can understand, plus why it's interesting. This is surprisingly difficult for founders who have perfect context about their own companies. Great communication requires replicating one's internal context in someone else's brain. Partner Tom wrote a skill teaching agents to condense company context into two-sentence descriptions. Partners recorded office hours sessions where founders practiced two-sentence descriptions with feedback, creating transcripts containing knowledge about effective description writing.
Feeding the meeting transcript back to the agent with instructions to improve the two-sentence description skill based on learned context made the skill noticeably better than human partners at writing these descriptions. This represents the micro-mechanism of organizational superintelligence: a prompt is written and used, artifacts like transcripts are generated, and these artifacts enable automated metaprompting that improves the skill daily. The skill becomes better than any individual human at that specific task.
The approach to building superintelligence inside a company involves composing everything that any given person can do, combining these capabilities in aggregate through particular processes. This creates a super organization. Every person can implement this at their own company and job. The reason to start a startup is that people become trapped in organizations run by powerful leaders who have resources and capital but do not believe this approach works.
Powerful organizations keep context locked down because they consider it unsafe. Building an AI native organization requires using AI as the building layer for everything rather than just a copilot. Organizations should start recording all artifacts including meeting recordings. Meeting recorders have gained popularity for coaching during meetings, but the value extends to improving all outputs including writing emails, communication, and planning by capturing the complete context.
Dario's essay identified that blockers to AI progression are not technical but social and cultural. Two years ago, recording meetings felt odd with unclear social etiquette around intrusiveness. Today, recording meetings has become default, especially on Zoom. When framed as a way for everyone in an organization to improve using the collective skill and instinct of colleagues, the approach becomes powerful.
A canonical two-sentence description skill serves not only to generate text snippets but helps people understand what makes for effective founder communication. This allows tapping into everything learned over years through conversations baked into the skill. This functions as a shared organizational brain, representing the closest approximation to connecting brains. Agents can conduct practice sessions and critique performance. Many possibilities emerge once knowledge exists in a place where agents can work with it.
Agent conversations are globally viewable by any full-time employee at YC by default. This decision required significant discussion about whether everyone seeing everything would be acceptable. The choice to keep conversations open proved valuable because people learned how to use the system by watching how others used it. Every agent conversation broadcasts internally to a Slack channel where anyone can join and learn.
Transparency solved multiple problems simultaneously. When one person began using the system heavily and creatively, others observed and discovered new use cases. Defaulting to public broadcast institutes social control on what people do with the system. This approach has proven reasonably effective in high-trust environments at keeping private information private while being more lenient on internal security. Agents are most powerful when given unrestricted access to context, which conflicts with how most organizations operate.
Truly agentic 1000x super intelligent organizations require being relatively egalitarian and trust by default. Neither trait characterizes most organizations. Founders must place these traits at the core of operations. This environment works best at startups where small groups operate with alignment in high-trust settings.
Organizations must be willing to spend $10,000 to $100,000 per year on tokens. Investing in skills while operating openly with teams creates a one-time time warp allowing leapfrogging of incumbents, Fortune 500 companies, and existing startups. Current spending of $100,000 or $1,000,000 annually will become commonplace within two years at reduced costs of $10,000, then a couple hundred dollars. Everyone will eventually adopt this approach as standard company practice.
The situation parallels companies purchasing computers for employees in the 1990s. These systems were expensive and only certain companies invested in them. Having a computer when competitors lacked them provided significant advantage.
This approach raises the floor by enabling new employees who previously required six months to ramp up to automatically acquire organizational context. They learn how star performers handle situations through AI apprenticeship rather than requiring partner time. The system simulates what it is like to be a specific person performing tasks like coaching founders on sales or giving specific advice. New entrants become mini versions of experienced team members faster.
New employees can ask agents questions they would be embarrassed to ask humans or hesitate to ask busy colleagues. This results in more questions being asked and answered, enabling faster ramp-up. After building agent infrastructure at YC, the essay "Horseless Carriages" was written.
The essay critiqued AI software being built by slotting small amounts of AI inside existing software rather than fundamentally rethinking architecture. The Gmail email writer exemplified keeping prompt context locked away from users under the assumption that developers should handle complexity and protect users from it. The potential for AI involves shifting control of software from developer to user.
As AI native software improves, it will resemble agents wrapping software deterministic tools rather than deterministic software wrapping AI. Chat serves as the interface because trust in agents increases over time, reducing the need for UI review of decisions. Agents can create just-in-time software, presenting specific views or building single-page JavaScript applications purposefully built for particular moments. These can exist as skill files callable anytime.
Chat functions as the better interface because it represents the closest thing to human language, and human language and writing constitute the closest expression of thinking. Chat serves as the closest stepping stone to clear intelligence. Constraining AI to very specific boxes limits capabilities. The interface supports multimodal input including text, voice memos, pictures, and files.
A half-million-line Rails application called Gary's List was built, including a full agentic framework for deep research and fact-checking. The application was constructed using 2013 web 2.0 approaches. Cloud Code enabled writing 40,000 lines of code in three days for Gbrain, Gary's List 2.0, which is fully open source. The agentic retrieval, voice extraction, and fact-checking components now exist inside Gbrain and were given to the Gary's List team as their own open source instance.
The monolithic writer chat interface contained bugs from reimplementing functionality already existing in other tools. The team now uses existing tools plus retrieval systems with extracted data through MCP, resulting in better performance. Half a million lines of Rails code represents rigid infrastructure that takes ten times longer to modify than dynamic alternatives. The same functionality can be achieved with 10,000 lines of TypeScript and 2,000 lines of markdown. Changes like adding biographies to specific paragraphs can occur without coding or complex eval infrastructure.
The best AI software tends to be very small, adding only the minimal code necessary to let the model perform. The ability to start with extremely simple systems requiring minimal understanding proves powerful. Most future software will follow this pattern. Open core succeeded by providing personality, persistence, memory, and duration. Claude Code emphasizes simplicity and minimal product size.
Pi represents a minimal coding agent harness used by technical teams. The software is self-extending and self-referential. OpenClaw was built on top of this system. Commercial software will increasingly ship with customization capabilities out of the box.
AI can function as either a centralizing or decentralizing force. The Gmail approach of preventing prompt changes exemplifies centralization. Two scenarios exist for the next 18 to 24 months, potentially extending to five years. The centralizing path involves five entities controlling advanced AI, compute, power, and data centers while preventing users from running their own prompts.
This scenario resembles the pre-personal computer era when computers cost hundreds of thousands to millions of dollars and remained tightly locked down by corporate policies. The computing revolution occurred when people gained access to personal computers for experimentation. A small priesthood and institutional base controlled the means of production.
The alternative path follows the Homebrew Computer Club model where Steve Jobs and Steve Wozniak soldered breadboards in a Mountain View garage and sold 500 Apple Ones. The current moment represents the Apple 1 stage where primitives are being developed. A billion users access ChatGPT with limited capabilities. MCP remains locked down, preventing easy database connections. Claude offers slightly more openness. Perplexity Computer provides the best current version but remains limited compared to OpenClaw and Hermes agent.
The revolution involves true personal AI where individuals run their own software, change their own prompts, test systems, maintain private repositories, and choose models including open-weight options. This represents the white pill for AI where a billion people control and program for themselves rather than having AI happen to them under corporate API control. The alternative to corporate control means AI becomes an extension of individual values and priorities rather than those of Meta, Alphabet, OpenAI, or Anthropic.
AI functions as empowerment rather than replacement. Technology development from mainframes to PCs to the internet represents a story of individual empowerment. AI eliminates drudgery that made jobs painful and enables more work than previously possible. Companies default to command and control rather than openness, with leadership potentially accessing tools while line-level staff cannot. Radically different organizations and different approaches to computing are necessary. These choices belong to the people building these systems.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
