Inside YC's AI Playbook

Chapter 1 · Narration: WOW Mentor

Inside YC's AI Playbook

What WOW Mentor takes apart here is the rare record of an organization running its own advice on itself. YC started from a small friction in its finance team and layered a single Postgres of shared context, a registry of more than 350 tools, and skills that reread their own transcripts nightly to grow better than the people who wrote them — making AI the building layer of the company rather than a co-pilot. Underneath sit three choices: a culture of egalitarianism and default trust that keeps context open instead of locked, a willingness to overspend on tokens now to live in the future early, and a philosophy of shifting control from developer to user. The conclusion WOW Mentor draws is that none of this is a rich company's magic trick — it is a door anyone can choose to open in their own organization today.

Today, WOW Mentor breaks down how Y Combinator — the firm famous for building other people's startups — actually uses AI inside its own walls. The conversation features Garry Tan and Pete Koomen, a partner who once built Optimizely and has since built all of YC's agent infrastructure. Pete explains that for years YC has mostly funded AI companies, and that the advice it kept handing those founders is now the same advice it runs on itself. What most people miss, he says, is that YC has quietly built the very tools it tells startups to build, becoming deeply AI-native itself. WOW Mentor picked this episode because it is rare to see an advisory firm experiment on its own body first, and that lab notebook holds a more honest strategy than any keynote. The story begins, improbably, in finance. About a year ago Pete and a few engineers started a project to help finance run its workflows — booking journal entries, logging priced rounds. He noticed a strange loop: finance would describe a procedure to the engineers; the engineers would build deterministic, purpose-built software that encapsulated it; then they'd hand it back, and around it went. What WOW Mentor highlights is that his complaint was not merely inefficiency but that control sat in the wrong place: the people who understood the workflow were in finance, yet only the engineers could turn it into software. His goal was to remove the engineers from the loop and let finance encode its own workflows — not in Ruby, but in English, with prompts. At the same moment, on his own laptop, Pete lived in a different world. Windsurf and Cursor were established, and Claude Code had just arrived; those tools gave him superpowers, and the gap with the old way of building software at YC grew wider. Narrow and cautious at the office, able to do anything at home — that split is what WOW Mentor treats as the root of the whole conversation — every experiment inside YC reduces to bringing the freedom he felt at home into the workplace. As later moments with Open Claw foreshadow, what held the world back was worry about security and privacy; worry a little less, and these things become unbelievably powerful. The first magic was not agentic coding but SQL. Garry Tan notes that years ago, when LLMs had arrived but agentic coding hadn't, one startup idea each of them had watched fail was precisely "get an LLM to write SQL." This time it worked — non-technical finance people began asking real questions with it. The purpose-built tool was rewritten into a general agent loop with a shared registry of YC-specific tools. Then a partner named Jared built two tools: one to run read-only SQL queries against the database, one to read the model files. Jared confesses he felt like he was breaking the rules; frustrated that narrow tools weren't powerful enough, he gave the agent complete access to the production database and pushed it out late one night. And it worked extremely well.

Why was a mere SQL query so powerful? What WOW Mentor emphasizes is a structural advantage. YC has always run mostly on its own software, all sitting on one Postgres database holding everything important — a table for every company it funded, a founders table, financial transactions, even the notes a partner leaves in the internal CRM. Functions most firms farm out to third-party SaaS, YC built itself, so the most important context lives in one place. Pete can now ask, "show me all the investors who invested in a space-related company in the last four batches," and an agent answers nearly arbitrary questions about the business. WOW Mentor's read: this weapon of the AI era accrued like interest to a firm that refused, long ago, to hand its control to outside vendors. But the real change wasn't that answers got easier — it exploded both the number and the daring of the questions. In the old BI-tool days, "which investors backed space companies?" meant hours of SQL, so unless it mattered you wouldn't bother. The name WOW Mentor attaches is the Jevons paradox: when the cost of getting something falls sharply, consumption surges rather than shrinks. The real leap is not the speed of answers but the widening of the horizon of questions — when things you'd never dare ask become routine, the total an organization can know about itself changes. Asked how an old-world organization grows wings, Garry Tan reaches for the history of data science. Just as Google's engineers invented Bigtable — one big, map-reducible table instead of schemas and joins — the same thing is happening again. He calls it a Karpathy-style knowledge LLM wiki and points to his own G brain: he pulls data from many systems into Open Claw and normalizes it into a schema that matters to him — really denormalization, pulling in every retrieval technique known, from RAG and graph RAG to reranking. Give the agent a soul and the data and let it know you, he says, and it sees around corners, returning answers only someone who knew you well could give. The rule WOW Mentor extracts: you can wrap data in an MCP, but for now a "Bigtable for the agent" is stronger. Pete diagnoses that we're still in the single-player era of agents. The popular harnesses — Claude Code, Codex, Pi, Open Claw, Hermes — are all built for one human on one machine, where they're staggeringly powerful. The unsolved problem is lifting that superpower to a team: the multiplayer harness. From that come two foundations he'd urge on any legacy organization. First, a common context layer — a warehouse where as much internal context as possible lives; like a coding agent in a monorepo, an agent on a single database with everything in one schema proved far stronger. Second, an internal tool registry. It began with about 20 tools and every team kept adding more until there are over 350 today; once tools for office hours, journal entries, and events all live in one place, they're available not only to internal agents but to Claude Code on individual machines — putting people and agents on the same plane of capability.

… The rest of this chapter continues in the full edition, available with a subscription.

Listen to this chapter