Y Combinator just open sourced the AI agent harness it uses to run its own company. It is called QM, and it is the clearest signal yet that 2026 is the year of the harness, not the model.
Here is everything you need to know about QM, why Y Combinator built it, and how it stacks up against the best agent harnesses available right now.
What is an agent harness?
An AI agent harness is the operating layer around a model. It decides what context the model sees, which tools it can call, where commands run, what persists between sessions, and when a human must approve an action.
The same model can perform differently depending on the harness around it. As one 2026 analysis put it, the frontier models have converged, so the wrapper now decides your experience.
Harnesses come in three flavors: terminal coding agents like Claude Code and Codex, personal assistants like OpenClaw, and now company wide systems like QM.
What is QM?
QM is a multiplayer agent harness for work, released by Y Combinator under an MIT license. It runs on Slack and the web, and it powers YC’s own accounting, legal, events, and engineering teams.
Most agents are personal assistants. QM is built for a whole company: each employee gets an isolated workspace with scoped memory, files, keychain, permissions, crons, and a durable sandbox, and teams can collaborate with the agent in channels and projects.
Y Combinator described it as easy to customize, like Hermes or OpenClaw, but useful for an entire organization.
Key features
QM ships with personal and shared scopes, so the agent works differently for each person and each team room. The same identity and configuration carries between Slack and the web app.
Admins get org level control: a security posture, which harnesses and models are available, and how skills are shared. Skills are scope owned and shareable by grant, with admin gated promotion to the whole org.
It can spin up internal web apps and publish them to the right people. Background crons and watches run work while nobody is watching.
On top of that, QM is genuinely multi vendor. Pi, OpenCode, Codex, and Claude Code can all drive the same core, so a deployment is never tied to one model company.
Security model
QM follows the same approach as local coding agents: the agent acts as the person it is working for, with their credentials and permissions, and everything it does is audited.
Orgs pick one of three postures: Strict pauses every tool call for human approval, Auto (the default) screens external data before it reaches the model, and Dangerous skips screening entirely.
Even in Dangerous mode, a predeclared command policy hard blocks things like recursive deletes and destructive SQL.
How it compares to the big names
| Feature | QM (YC) | OpenClaw | Claude Code | Pi | Hermes Agent |
|---|---|---|---|---|---|
| Company-wide deployment | Yes, built for orgs | No, personal | Partial, via teams | No, personal | No, personal |
| Slack / chat interface | Yes | Yes | Yes | No | Yes |
| Per-user sandbox | Yes, durable | Yes | Yes | Yes | Yes |
| Skills / plugins | Yes, shared by grant | Yes | Yes | Yes | Yes |
| Crons / scheduling | Yes | Yes | Yes | No | Yes |
| Internal web apps | Yes | No | No | No | Yes |
| Multi-model / multi-harness | Yes (Pi, Codex, OpenCode, Claude Code) | Yes | Anthropic-first | Yes | Yes |
| License | MIT, open source | Open source | Closed | Open source | Open source |
The short version: QM is the only one designed from day one for company wide use with per employee isolation. OpenClaw is the best personal always on assistant. Claude Code is the reference terminal coding agent.
The adoption picture
OpenClaw leads in raw GitHub adoption with roughly a quarter million stars. Claude Code sits around 120,000. QM, being one day old, started at about 3,900 stars and is climbing fast.
Star counts measure hype, not capability. QM’s real validation is internal: Y Combinator runs its own accounting, legal, events, and engineering workflows on it.
Who should use what
Startups and companies that want a self hosted company agent should look at QM. It is the only option in this list that gives every employee their own isolated workspace with org level admin control.
Individuals who want an always on assistant across messaging apps should pick OpenClaw. It connects to Telegram, WhatsApp, and other channels, and it has the biggest community.
Developers doing repository work should use Claude Code, Codex, or Pi depending on their model preference. Pi is the open source, model agnostic choice.
Builders who want a configurable agent with provider choice, persistent learning, messaging, and scheduling should evaluate Hermes Agent, which QM itself cites as an inspiration for customizability.
The bottom line
Y Combinator open sourcing QM is a landmark moment for the harness ecosystem. It takes the multi agent company concept from startup folklore to a real, MIT licensed reference implementation.
The takeaway for developers: stop picking models and start picking harnesses. The model decides raw intelligence, but the harness decides whether that intelligence is safe, reliable, and actually useful to a team.
QM is the newest player, but it is the only one that treats the whole company as the unit of design. That is a bet worth watching.
