Source: MarkTechPost — 2026-08-03
Summary
Y Combinator open-sourced QM, the multi-agent harness it uses to run its own internal operations — accounting, legal, events, and even the engineering work of building QM itself — under the MIT license. The core is a headless TypeScript/Node service on Fastify with Postgres for session, memory, and queue state, and it's model/harness-agnostic: Pi, OpenCode, Codex, and Claude Code can all drive the same underlying core through Slack and web plugins plus an admin panel. Security is configurable per-deployment on a spectrum from "strict" (human approval on every tool call) to "auto" (automated content screening) to "dangerous" (no gating at all), with skills scoped per-person or per-room and admin-gated before they can be promoted org-wide.
Key Takeaways
- QM is dogfooded infrastructure, not a demo — YC runs real company functions (accounting, legal, events) through it, which is a stronger signal of production-readiness than a typical framework release.
- Architecture is deliberately boring and operational: headless TypeScript/Node on Fastify, Postgres for session/memory/queue state — built to be run and debugged like a normal backend service, not a research prototype.
- It's explicitly harness/model-agnostic, letting Pi, OpenCode, Codex, and Claude Code all drive the same core — a bet on QM as the durable coordination layer while the model/harness layer churns underneath it.
- Sandboxing is scoped per-person and per-room (files, keychains, permissions), which is the mechanism that makes "multiplayer" agents safe to run without one agent's session bleeding into another's credentials or files.
- The security dial — strict/auto/dangerous — lets an org trade approval friction for speed deployment by deployment, rather than picking one global policy for every agent action.
- YC's stated target profile is a 10-500 person company with at least one platform engineer, meaning this is aimed at teams that can operate infrastructure, not a plug-and-play consumer tool.
Reel Script
Hook (18s)
Y Combinator just open-sourced the exact system it uses to run its own accounting, legal, and events — built and maintained by AI agents. And they gave the whole thing away for free.
Core Concept (80s)
QM is what's called a multi-agent harness — think of it as the operating system that lets multiple AI agents share the same company context instead of each one working in its own isolated bubble. The core is a headless backend: a TypeScript and Node service running on Fastify, with Postgres holding session state, memory, and task queues. What makes it interesting is that it's model-agnostic — you can plug in Pi, OpenCode, Codex, or Claude Code, and they all drive the same underlying core through the same Slack and web interface. That's a real architectural bet: instead of building the product around one specific model, YC built a coordination layer that outlives whichever model or coding agent is hot this quarter. And this isn't a demo repo — YC actually runs its own company operations through it, including the engineering work of building QM itself, which is about as strong a "we trust this in production" signal as you can give.
Hands-On (60s)
The part worth sketching out is the security model, because it's the piece that decides whether you'd actually trust this in your own company. It's a three-position dial, set per deployment. "Strict" means a human has to approve every single tool call the agent wants to make — safest, slowest. "Auto" swaps that human gate for automated content screening — the system checks the action itself rather than waiting on a person. "Dangerous" removes the gating entirely and just lets the agent run. Layered on top of that, every skill an agent can use is scoped to a specific person or room by default, and only an admin can promote a skill to be available org-wide. That combination — a tunable approval dial plus scoped-by-default permissions — is the actual answer to "how do you let agents touch real company systems without it going sideways."
Takeaway (25s)
My honest take: this is the first agent harness release that reads like real ops infrastructure instead of a hackathon wrapper, precisely because YC is betting its own back office on it. If you're a platform engineer at a 10-to-500-person company, QM is worth a serious eval this week — go pull the repo.