{"slug": "y-combinator-open-sources-qm-the-ai-agent-harness-it-uses-to-run-itself", "title": "Y Combinator Open-Sources QM, the AI Agent Harness It Uses to Run Itself", "summary": "Y Combinator has open-sourced QM, an AI agent harness it built to manage fleets of internal agents, under the MIT license on GitHub. The harness, designed for Slack and the web, provides scoped memory, permissions, and sandboxes for each user and room, and supports multiple AI models including Pi, OpenCode, Codex, and Claude Code. YC says it provisioned more than 50 Hermes agents before developing QM, which it describes as a 'quartermaster' coordinating the operating layer around agents.", "body_md": "*Y Combinator has released QM, the agent harness it built after running fleets of internal agents, and the useful part is simple: you can inspect the system instead of trusting a demo.*\n\nY Combinator has put QM on GitHub, and the timing is not subtle. Startups are already trying to turn AI agents from clever side tools into something closer to company infrastructure. YC's answer is to release the harness it says lets startups, and YC itself, work with a fleet of agents across Slack and the web.\n\nThe repository is live under the yc-software organization, and GitHub lists it under the MIT license. That's the practical fact at the center of the story. You can fork it. You can modify it. You can run it commercially. YC isn't asking founders to join a waitlist before they see how the thing works.\n\nQM is not just another chatbot wrapper. The GitHub README describes it as a multiplayer agent harness for work, built for Slack and the web, with a design that gives each person and each room its own scoped memory, files, keychain view, permissions, crons, web apps, and durable sandbox. That is the part founders should pay attention to first.\n\nA single assistant can be useful. A company assistant is harder. If a finance employee and a founder both use the same agent system as the engineer who built it, context and credentials become the whole problem. QM's answer is not to pretend that problem disappears. It gives each scope its own boundary.\n\n## The design starts with isolation\n\nAccording to YC's QM page, the accelerator had already experimented with several harnesses before releasing this one. It started with a basic agent loop in Ruby that could access internal data, then extended that system with features like crons and webhook triggers. Later, YC says it provisioned more than 50 Hermes agents for individual employees to use as personal assistants.\n\nThat is a useful admission. Managing a fleet even at that size became difficult, YC says, and OpenClaw pushed the team in a new direction. The result is QM, short for quartermaster, a name YC defines in a note on the project page as the person on a ship who coordinates belowdecks to keep things in order.\n\nThat metaphor is doing real work here. QM is less about one dazzling model answer and more about keeping the operating layer around agents in order: identity, policy, scheduling, persistent state, tool access, and the sandbox where commands actually run. The README says the core runs TypeScript on Node, uses Fastify for HTTP, puts session history and memory in Postgres, and offers Slack as an optional in-process plugin.\n\nDry details matter. They tell you whether a project is meant to be admired or installed.\n\n## No favorite model\n\nThe sharpest product decision in QM is that YC did not tie it to one model or one coding tool. The README says Pi, OpenCode, Codex, and Claude Code can all drive the same core. That matters if you're building company workflows today, because nobody serious should assume the best agent stack in August 2026 will still be the best one six months from now.\n\nVendor lock-in is not an abstract fear for startups. It shows up when your internal tools depend on one model's quirks or one hosted platform's pricing - the memory format gets locked in too. QM's pitch is that the substrate should sit behind interfaces, so a deployment can swap models, harnesses, session stores, sandboxes, and memory implementations without rebuilding the company around a single vendor's roadmap.\n\nHere's the thing: free software is not automatically useful software. A GitHub repo can be open and still be a weekend experiment. QM has more weight because YC published the operational scars next to the code. The project page says YC wanted something as flexible as Hermes but with the simplicity of its original system, and also wanted something it could own and host itself.\n\nThat is exactly the pressure many startups now feel. They want agents inside Slack, in repositories, around internal documents, and eventually inside repeatable workflows. They also want control over credentials, logs, permissions, and the cloud account where the system runs. You cannot hand-wave that part away once an agent can touch financial data or read your code.\n\nThe repository's security section is unusually plain about the tradeoff. QM follows the pattern of local coding agents, where the agent acts as the person it's working for, using that person's credentials and permissions, while its actions are audited. It also offers security postures ranging from strict human approval to a dangerous mode with no pauses between tool calls. Frankly, naming the dangerous mode is better than pretending it isn't dangerous.\n\nFor founders, the real test is not whether QM sounds impressive in a launch post. That test is simple. Can your own team read the getting-started guide and actually decide where an agent should be allowed to act? The code is at github.com/yc-software/qm, and YC's own project page points readers there directly.\n\n**Also read:** [OpenAI's and Anthropic's AI Agents Escaped Testing and Hacked Real Firms](https://startupfortune.com/openais-and-anthropics-ai-agents-escaped-testing-and-hacked-real-firms/); [Reddit's Stock Sinks After Earnings as Google's AI Overviews Eat Its Traffic](https://startupfortune.com/reddits-stock-sinks-after-earnings-as-googles-ai-overviews-eat-its-traffic/); [Google pulled its Earth AI image tool in under 24 hours after users faked disasters and military strikes](https://startupfortune.com/google-pulled-its-earth-ai-image-tool-in-under-24-hours-after-users-faked-disasters-and-military-strikes/)", "url": "https://wpnews.pro/news/y-combinator-open-sources-qm-the-ai-agent-harness-it-uses-to-run-itself", "canonical_source": "https://startupfortune.com/y-combinator-open-sources-qm-the-ai-agent-harness-it-uses-to-run-itself/", "published_at": "2026-08-01 05:06:45+00:00", "updated_at": "2026-08-01 05:25:58.703631+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "ai-infrastructure", "developer-tools"], "entities": ["Y Combinator", "QM", "GitHub", "OpenClaw", "Hermes", "Pi", "OpenCode", "Codex"], "alternates": {"html": "https://wpnews.pro/news/y-combinator-open-sources-qm-the-ai-agent-harness-it-uses-to-run-itself", "markdown": "https://wpnews.pro/news/y-combinator-open-sources-qm-the-ai-agent-harness-it-uses-to-run-itself.md", "text": "https://wpnews.pro/news/y-combinator-open-sources-qm-the-ai-agent-harness-it-uses-to-run-itself.txt", "jsonld": "https://wpnews.pro/news/y-combinator-open-sources-qm-the-ai-agent-harness-it-uses-to-run-itself.jsonld"}}