YC Open-Sourced the AI System Running Its Company

Y Combinator just released the internal AI agent system it used to run accounting, legal, and engineering — and to build itself.

The world's most famous startup school showed its hand

Y Combinator — the accelerator behind Airbnb, Dropbox, and hundreds of others — has been quietly running its own company with an AI agent system called QM (short for Quartermaster). Accounting, legal, events, engineering. Not demos. Not experiments. Their actual operations.

Last week they put the whole thing online for anyone to use.

What makes QM different from the AI assistants you've probably tried is that it's built for a whole team, not a single person. Every employee gets their own memory, their own files, their own permissions. So your accounts person and your designer can both be working with AI agents at the same time without one accidentally seeing or touching the other's stuff. It's a bit like giving each department their own well-briefed assistant that also happens to know what the rest of the company is doing — but only as much as it should.

It also doesn't force you to pick a favourite AI model. Claude, OpenAI's Codex, and others can all drive it. That matters because the AI landscape is still shifting fast, and being locked to one provider today is a real risk.

Real teams are already using it to auto-fix broken code pipelines overnight, generate morning reports, and flag problems the moment they appear in production.

For a business owner, the headline isn't the technology — it's the signal. When a top-tier investor runs their own company on AI agents and then hands you the blueprint, something has genuinely changed.

Words worth knowing

AI agent — an AI that doesn't just answer questions but takes actions: sending emails, reading files, checking systems, fixing problems. Think less chatbot, more capable intern.

Multi-agent — several AI agents working at once, each with a specific job, coordinated like a small team rather than a single helper.

Open-source — the code is published publicly for anyone to inspect, use, or build on. No licence fee, no black box.

Model-agnostic — works with different AI providers (Claude, OpenAI, etc.) so you're not stuck if a better option appears later.

Worth asking yourself: which part of your week is the most repetitive, the most rules-based, the most "I could write this down as a checklist"? That's usually where agents start earning their keep.