Y Combinator open-sourced the multi-agent tool it built to run itself

Y Combinator open-sourced QM, the multi-agent harness it built for its own internal use, on July 31, 2026. The accelerator’s own announcement describes it as a “multiplayer agent harness” meant for a whole company rather than one person — YC says it already runs QM across its accounting, legal, events and engineering functions, including using it to build QM itself. The project is released under an [MIT license](#/dictionary) on GitHub, where it drew roughly 1,900 stars within hours of the announcement.

The design choice that separates QM from a personal AI assistant is scope: every employee and every project channel gets an isolated workspace with its own memory, files, permissions and scheduled automations, while shared “skills” can be promoted from one person’s workspace to the whole organization by an admin. It ships with three configurable security postures — Strict, where every tool call the agent wants to make pauses for a human to approve except the two that don’t change anything; Auto, the default, where a classifier screens external data before it reaches the model; and Dangerous, with no screening or approval pauses at all. Command policies the organization predeclares are enforced regardless of which posture is active.

QM doesn’t lock a company into one model vendor’s agent product. It supports Pi, OpenCode, [OpenAI’s](#/company/openai) Codex and [Anthropic’s](#/company/anthropic) Claude Code as interchangeable underlying harnesses, and organizations deploy it into their own cloud accounts on Fly or AWS rather than a YC-run service — the `qm` CLI generates a deployment repository, with no production CI included by default. That combination — open license, vendor-neutral model support, self-hosted deployment — is a deliberate contrast with the closed, single-vendor agent products major labs have been shipping this year.