Inside YC's AI Playbook: Pete Koomen on Building Agent Infrastructure at Y Combinator

Pete Koomen, general partner at Y Combinator and founder of Optimizely, describes how he led the construction of YC's internal agent infrastructure from scratch over the past year — starting with a finance-team pain point and snowballing into a 350-tool shared system that has transformed how the organization operates. He argues that the key to building superintelligence inside a company is not using AI as a copilot feature, but as the foundational operating layer, paired with a trust-by-default culture and radical context sharing.

Key Takeaways

A single Postgres database containing all organizational context, combined with read-only SQL access for agents, was the first transformative unlock — dramatically increasing both the number and complexity of questions people would ask 07:37
A shared tool registry (now 350+ tools) is what turns generic agents into something useful at work; every team at YC contributes domain-specific tools 14:49
Skillify + DRY/MECE resolver discipline keeps the tool/skill space non-redundant and collectively exhaustive, which Koomen argues is the optimal resolver design 16:24
A nightly self-improving dream cycle reads all agent conversation transcripts and automatically improves skills — the two-sentence pitch skill became "better than any of us individually" after this loop ran 18:23
Making all agent conversations publicly viewable internally solved the adoption and security problems simultaneously — people learned by watching peers, and social visibility served as a soft access control 27:43
Truly superintelligent organizations require being egalitarian and trust-by-default; neither trait is the default in most companies 29:18
Koomen predicts AI will follow the PC arc: individual empowerment rather than replacement, but only if deliberate choices are made against centralization 43:06

The Finance-Team Origin and the SQL Unlock

Koomen traces the project's origin to frustration with the classic loop: finance describes workflows → engineers encode them deterministically → hand back → repeat 03:40. Around the same time, tools like Windsurf, Cursor, and Claude Code were emerging, and the contrast between his personal machine capabilities and the narrow tools available at work felt like a "bigger and bigger divide" 04:33.

The decisive breakthrough came when engineer Jared built two tools: read-only SQL access to YC's production Postgres database and the ability to read model files 06:21. Koomen acknowledges this felt like "breaking the rules" — granting near-unrestricted production DB access — but it worked 06:56. YC's advantage was structural: because it has always run on its own software, all organizational context (companies, founders, financials, CRM notes) lives in one schema 07:52. This enabled arbitrary natural-language queries across the entire business, illustrating Jevons paradox: lowering the cost of asking questions dramatically increases the volume and ambition of questions asked 09:14.

The Shared Tool Registry and Skill Architecture

From ~20 initial tools, the registry grew to 350+ as every team added domain-specific capabilities: office hours management, journal entry booking, event management 14:40. Koomen stresses that a shared, centralized tool registry is what enables both internal YC agents and individual Claude Code sessions on personal machines to draw from the same capabilities 15:15.

Skills are an abstraction layer over tools. Koomen describes a progression — handwritten prompts → written skills → meta-prompting to auto-improve skills — that he observes being independently rediscovered across Claude Code, Codex, Open Claw, and Hermes Agent 18:05. He advocates for a "check resolvable" meta-skill that enforces DRY (don't repeat yourself) and MECE (mutually exclusive, collectively exhaustive) discipline across the resolver table, arguing that one parameterized tool beats ten overlapping ones 16:36.

The Self-Improving Dream Cycle and Organizational Brain

YC runs a nightly agent that reads all employee agent conversation transcripts, identifies missed context and improvement opportunities, and updates skills automatically 19:28. This is structurally identical to what Koomen calls the "dream cycle" in Open Claw and G Brain 19:42.

The concrete example: YC partner Tom wrote a skill for generating two-sentence pitches (concise founder descriptions — "what is it, and why is it interesting?" 20:26). Other partners ran a group office hours session coaching founders on the skill; the meeting transcript was fed back to the agent with an instruction to improve the skill. The result noticeably outperformed any individual partner 22:57. Koomen's conclusion: "This is how superintelligence happens inside organizations" 23:04 — encode a practice, accumulate transcripts as training artifacts, auto-improve via meta-prompting, and repeat across every organizational skill.

Horseless Carriages and the Case for Agent-First Software

Koomen's essay "Horseless Carriages" critiques AI products that slot AI as a small feature inside a lot of deterministic software — exemplified by Gmail's AI email writer with a locked system prompt the user cannot modify 32:47. He argues the correct inversion is: agents wrapping deterministic tools, not deterministic software wrapping an AI 34:06.

He advocates for just-in-time software: minimal, small codebases that let the model shine, with complexity added on-demand rather than pre-built 38:57. The best AI software he has used "tends to be very small" and adds "the smallest amount of code ahead of time that you need in order to let the model shine" 38:57. He observes that self-extending minimal harnesses like Pi (the open-source coding agent underlying Open Claw) exemplify this — Pi can modify and extend itself 40:09.

On interfaces, Koomen changed his mind: chat is the right default interface because it is the closest approximation to human expression of thinking, and increasingly users can trust agents enough to not need heavy UI review layers 34:22.

Centralization vs. Decentralization: The Personal AI Stakes

Koomen frames the coming 18–24 months as a fork between two futures 41:07:

Centralized control — a handful of large players lock down prompts, restrict DB access, and deliver AI that "happens to you" rather than is controlled by you; analogous to a world where personal computers never displaced mainframes 41:16
Personal AI revolution — open, user-controlled systems where individuals can modify prompts, choose models (including open-weight), and connect their own data; analogous to the Homebrew Computer Club and the Apple I moment 42:51

Koomen argues "we are at the Apple One moment right now" 43:07 — primitives are being discovered, but packaging and mass access remain unsolved. He is explicit that he wants "a billion people to actually control and program for themselves" what these systems do 44:37, and warns that achieving this requires deliberate organizational and product choices against the default of command-and-control, locked-down AI deployment.