Y Combinator Open-Sources QM: A Multiplayer AI Agent Harness
Y Combinator has released QM under an MIT license, providing an enterprise-grade multiplayer AI agent harness that runs in Slack and the web.
Stock photo for illustration only, not from the actual event
- QM is enterprise organization software rather than a desktop application.
- Each user and chat room gets an isolated workspace, memory, and sandbox.
- It is harness-agnostic, supporting Pi, OpenCode, Codex, and Claude Code.
- The project is available under an MIT license for immediate cloud deployment.
Y Combinator has open-sourced QM, an MIT-licensed multiplayer AI agent harness designed specifically for organizational use. The software is ready for deployment today, functioning as organization software rather than a desktop application. A successful setup requires a cloud account, a Postgres database, and someone comfortable with managing infrastructure. The ideal fit is a startup or mid-sized company with roughly 10 to 500 employees and at least one platform engineer, while larger enterprises can run it after reviewing their security posture.
Deploying QM does not require cloning the entire repository. Instead, an organization-owned deployment repository depends on @yc-software/qm and runs qm init with an organizational slug alongside a Fly.io or AWS target. This initialization process creates a deployment skill that guides the administrator through infrastructure setup, web sign-in, connectors, optional Slack access, and live verification. Sign-in defaults to a built-in broker that emails a one-time link, though an external identity provider can be integrated. Every deployment operates strictly within the operator's own cloud account.
Stock photo for illustration only, not from the actual event
Most conventional AI agents are built as personal assistants, but the YC team notes that scaling a single assistant across an entire enterprise quickly introduces complexity. QM addresses this by assigning each employee an isolated workspace that others cannot interfere with, while allowing seamless collaboration within channels, group messages, and projects. Every individual and room maintains its own scoped memory, files, keychain view, permissions, cron jobs, web applications, and durable sandbox, with identical identity and configuration carrying over between Slack and the web application.
Y Combinator's release of QM highlights a major shift toward collaborative, multi-agent enterprise architectures where AI tools are deeply integrated into team workflows. By remaining harness-agnostic and giving organizations granular control over execution postures and data screening, QM bridges the gap between powerful developer tooling and strict corporate security requirements.
At its core architecture, every turn routes through a central headless core that handles APIs, identity, policies, and scheduling while driving the agent loop. A Postgres layer securely maintains user data, session history, and durable states. The agent relies on a fixed, lightweight tool surface, including an execute tool that runs commands inside the scope's dedicated isolated sandbox. Notably, QM is harness-agnostic, meaning Pi, OpenCode, Codex, and Claude Code can all drive the exact same core without tying a deployment to a single vendor.
QM also incorporates robust security postures that organizations can configure based on their risk tolerance. The Strict posture pauses for human approval on every harness tool call, while the default Auto posture screens external data and tool results using a classifier. The Dangerous posture removes content screening while retaining hard-coded command policies that block destructive actions such as recursive deletes and dangerous SQL queries. Comprehensive security limitations and threat models are documented within the official repository.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment