Y Combinator open-sources QM, the harness running its company

Y Combinator open-sourced QM on July 29, the agent harness it runs its own company on. The repo hit 8,420 stars and 887 forks in five days, per GitHub.

That matters for any team betting on a single agent CLI. QM argues the real cost of switching tools was never the model, it's the memory, permissions, and cron jobs stuck in whichever harness you picked.

The design choice worth studying isn't the feature list. It's that QM keeps state in Postgres and treats the model and the CLI as swappable parts.

Pi, OpenCode, Codex, and Claude Code all attach to the same core. Switching which agent tool you run doesn't mean losing your crons, keychains, or accumulated context.

Three things landed in the same four days, all pointing the same direction. The MCP spec, updated July 28, dropped protocol-level sessions in favor of server-minted handles. A directory called best-of-Agent-Harnesses started cataloging 141 harnesses, complete with its own recommend-and-pick tool.

AgentSky took the top Product Hunt spot on August 3 selling the same idea as a product. Pick a harness and a model, launch once, and reach it from Slack, WhatsApp, or the CLI with history intact.

Steve Yegge disagrees. He argues harnesses get bespoke to each app rather than shared.

He backs it with a claim of 175 to 250 commits a day across 12 Max accounts plus his own. I think he's half right. The harness code goes bespoke, but the state underneath it becomes shared infrastructure, and QM is the clearest bet yet on that split.

Clone the repo and read the permission model, even if you never run it.