Y Combinator open-sources QM multi-agent harness
Y Combinator has open-sourced QM, its internal multi-agent harness, under the MIT license, giving startups a vendor-agnostic framework to deploy collaborative AI agents in Slack and the web.

Y Combinator has released QM, or quartermaster, its internal multi-agent harness, as an open-source project under the MIT license. Described by its creators as an early and experimental tool that still contains bugs, QM is designed to run collaborative agent workflows across Slack and web interfaces. Y Combinator already uses the system internally to manage operations in accounting, legal, events, and engineering. Rather than acting as a simple desktop application, QM is enterprise-grade organizational software meant to be deployed within a company's own cloud infrastructure.
The platform is optimized for startups and mid-sized businesses ranging from 10 to 500 employees, provided they have at least one platform engineer to manage the setup. Deployment does not require cloning the main repository; instead, organizations can use the @yc-software/qm package and run an initialization command targeting either Fly.io or AWS. The system relies on a PostgreSQL database to manage session history and user data. It is highly flexible regarding underlying models, allowing practitioners to drive the central headless core using various engines such as Pi, OpenCode, Codex, or Claude Code without vendor lock-in.
For practitioners, QM shifts the paradigm from single-user personal assistants to isolated, collaborative workspaces. Every user and room receives its own scoped memory, files, keychain view, permissions, and durable sandbox. The core runs TypeScript directly on Node and uses Fastify for HTTP, while the web UI is built with Vite and Lit, and the Slack integration uses Bolt. To ensure safety, the harness implements three distinct security postures: Strict, which pauses for human approval on every tool call except two no-effect turn enders; Auto, which uses a classifier to screen incoming data; and Dangerous, which removes pauses and screening. Crucially, a hard-coded command policy blocks destructive actions like recursive deletes and destructive SQL queries across all three postures.
This is our own summary of reporting by MarkTechPost



