YC's open-source qm framework gives every teammate a scoped AI agent workspace with shared memory, permissions, and full audit trails.
Most AI agent tools so far have been built around a single user talking to a single assistant. Y Combinator just open-sourced something built for a different reality: a company where every employee has their own AI agent, all of them working inside shared channels, shared projects, and shared accountability. The project is called qm — a "multiplayer agent harness for work" — and it landed on Hacker News' front page with over 600 points, a strong signal that this is resonating well beyond YC's own portfolio.
For agencies building automation and AI workflows for clients, qm is worth a close look, because it's less a chatbot and more a blueprint for how AI agents should actually be deployed inside a real organization.
qm gives each person in a company their own isolated agent workspace — scoped memory, files, a keychain view, permissions, scheduled jobs (crons), even web apps — all separate from everyone else's. But it's not purely individual: people and their agents can also collaborate together inside channels, group messages, and shared projects, the same way a human team would work in Slack.
Critically, the agent acts as the person it's working for — using their credentials and permissions — with every action logged and auditable. That's a meaningful design choice: it means an agent's mistakes and access are bounded by the same limits as the human it represents, not by some separate all-powerful service account.
qm ships with three built-in security postures, and the choice between them says a lot about how seriously the project takes agent risk:
This tiered approach matters for agencies because it gives you a real answer to the question every client eventually asks: "what stops the agent from doing something it shouldn't?" Instead of a vague promise, qm gives you a configurable, auditable policy layer.
qm isn't tied to one AI vendor. The same core architecture is designed to work with Claude Code, Codex, OpenCode, and Pi, so teams aren't locked into a single model provider's roadmap or pricing. That flexibility is increasingly important as model pricing and capabilities shift month to month — an agency building infrastructure today doesn't want to re-architect it if the best model changes in six months.
The whole project is released under an MIT license, meaning it can be self-hosted, forked, and customized freely — no seat licensing, no vendor gate.
Most client-facing AI automation work eventually runs into the same wall: a single assistant can only go so far before a business needs multiple agents doing different jobs, working within different permission boundaries, and leaving an audit trail a client's IT or compliance team can actually review. qm is one of the first credible, open-source attempts to solve that at the "whole company" level rather than the "one user, one bot" level:
It's early — this is a fresh open-source release, not a mature, battle-tested platform — but the architecture reflects genuine thinking about how AI agents should be deployed at company scale, which makes it a strong one to evaluate before building similar infrastructure from scratch.
Our cutting-edge features simplify collaboration and creativity, making your workflow intuitive and efficient. Transform your vision into reality effortlessly with Hadidiz Flow.
