AQ vs qm: a company-wide agent harness vs a multiplayer coding workbench

Published September 1, 2026 · by the AQ team

Choose qm if you want an open-source agent layer for your whole company (accounting and legal as much as engineering), you are comfortable deploying and operating it in your own cloud, and chat is how you want people to work with agents. Choose AQ if the job is specifically engineering: a team running coding agents like Claude Code and Codex where everyone can watch the actual live terminal session, edit the code, and click through a running preview. The one-line difference: qm is company software you assemble and operate; AQ is a finished coding workbench you sign into.

The two products share a category word and little else. Y Combinator open-sourced qm (short for Quartermaster) on July 31, 2026 under the MIT license, describing it as "a multiplayer agent harness for work. In Slack and on the web." AQ uses the same framing (the multiplayer coding harness) but aims it at one audience: engineering teams shipping code with agents. Everything below about qm is verified against its own repository as of September 2026.

AQ vs qm at a glance

What is qm?

qm is the agent harness Y Combinator built for itself and then open-sourced. The design center is the company, not the individual: instead of one personal assistant, every employee and every shared room gets an isolated scope with its own memory, files, keychain view, permissions, scheduled jobs, web apps, and a durable sandbox where an execute tool runs commands. People work with agents in Slack or in qm's web UI, independently in their own scope or together in channels and projects.

Several design decisions stand out as of September 2026. The core is harness-agnostic: Pi, OpenCode, Codex, and Claude Code can all drive it, so a deployment is not tied to one vendor. Skills are owned by a scope, shareable to other scopes by explicit grant, and an admin can promote a skill to the whole organization. Background work is first-class: crons, watches, and webhooks keep agents working when nobody is in the room. For engineering specifically, the README describes agents working inside existing repositories: running tests, opening pull requests, monitoring CI, and checking logs. And security is a configurable posture: Strict requires human approval on every tool call, Auto (the default) puts a classifier in front of external content before it reaches the model, and Dangerous turns screening off.

Equally notable is what qm says it is not. It deploys from an npm package (version 0.1.5 as of late August 2026) into an organization-owned deployment on Fly.io or AWS via a qm init command, assumes Postgres and someone comfortable with infrastructure, and its own security documentation is explicit that it is not a hardened public or multi-tenant service boundary: it assumes one organization of authenticated internal users. The same document notes that org admins can read scoped content without additional user approval (audited), and that credentials materialized inside a sandbox are readable by processes in that sandbox.

What is AQ?

AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud.

Agents run as the real CLIs your engineers already use (Claude Code, Codex, Cursor Agent, Kimi, Grok, or plain shells) in persistent tmux sessions on the team's VM, streamed live to the browser. Sessions survive a closed laptop and resume from any device, and teammates open the same workspace and watch the same live session. Each workspace is an isolated git worktree (branch ai/{id}-{slug}, dependencies installed automatically, one-click rebase onto main); agents commit, push, and open PRs under per-user GitHub auth, and AQ tracks the PRs per workspace. Every workspace gets a live dev-server preview with share links that need no account to view, and comments pinned on the preview can be sent to the agent as prompts. Labeling a Linear issue "ai-task" creates a workspace with two-way status sync. Each user signs into the CLIs with their own Claude or OpenAI account, and AQ never marks up model usage.

The core difference: assemble a platform vs open a workbench

qm answers the question "how do we give the whole company an agent layer we control?" The answer is software: you deploy it into your own cloud, wire up Slack and sign-in, decide postures and skill grants, and operate it from version 0.1.x onward. In exchange you get something AQ does not try to be: one agent surface for legal, accounting, events, and engineering, fully yours to customize.

AQ answers a narrower question: "where does the engineering team's agent work live?" Because the scope is code, the product is built around what coding work needs: the real CLI session visible live (not a chat rendering of it), a code editor next to the terminal, a running preview of the app, worktree isolation so parallel work does not collide, and PRs tracked where the work happened. Typing into someone else's terminal is delegated, not ambient: the session owner approves a control request in one click and keeps their own input.

Put differently: in qm, a coding agent's run is one kind of work happening inside a company platform you operate. In AQ, the coding session is the product, and the whole team can stand around it. The two do not compete for the same seat as much as the shared word suggests. For the category question the two names raise, see harness of harnesses.

Where qm wins

• Whole-company scope. qm is built for every department, not just engineering. If the goal is agents for legal, finance, and ops under one roof, AQ does not address that.
• Open source and free. MIT-licensed, self-deployed, fully forkable. You can read every line, customize anything, and pay only your own cloud costs.
• Slack-native. For teams that live in Slack, agents in channels and DMs with unified identity is a natural fit.
• Configurable security postures. Strict, Auto, and Dangerous let an admin pick the approval and screening level per deployment, and skill promotion is admin-gated.
• Background automation primitives. Crons, watches, and webhooks are first-class, so unattended recurring agent work is a designed-for pattern.

Where AQ wins

• The live session, not a chat about it. AQ streams the actual CLI terminal to everyone in the workspace. Watching, steering, and reviewing what an agent really did is the product, not a transcript rendered into chat.
• A finished product, operated for you. No deployment repository, no Postgres, no 0.1.x software to run. The Free sandbox gives one person a private machine in an isolated network with nothing to install and no time limit.
• Built for code end to end. Isolated git worktrees per workspace, auto dependency install, one-click rebase onto main, per-user GitHub auth, PRs tracked per workspace, and live previews where pinned comments become prompts to the agent.
• Issue tracker intake. Label a Linear issue "ai-task" and a workspace appears in the team sidebar, with status syncing both ways and ownership following the assignee.
• Per-user CLI logins. Each engineer's Claude Code or Codex runs under their own subscription, and AQ never marks up model usage.
• No shared multi-tenant execution tier. Work runs on a VM you connect from your own cloud or a dedicated always-on AQ-managed VM in its own isolated network.

Pricing compared

qm is free software: MIT-licensed, with no vendor to pay. The real costs are your own: the Fly.io or AWS resources it runs on, Postgres, model access for the harnesses you connect, and the engineering time to deploy, upgrade, and operate an early-stage codebase. For a company that wants agents everywhere and has platform engineers to spare, that trade can be excellent.

AQ is per-seat: the Team plan is $50 per user per month in early access (standard $200), billed monthly, covering VMs you connect from your own cloud or a dedicated always-on AQ-managed VM in its own isolated network, with the rate locked for your first 12 months. Model usage stays on each user's own Claude or OpenAI subscription with no markup. The Free plan is a personal sandbox with no time limit, and inviting a teammate costs nothing until the person accepts and joins.

Bottom line

qm is a serious piece of open-source software with a clear thesis: the useful unit of an agent harness is the company, and a startup with infrastructure skills should own that layer itself. If that is your project, qm is the strongest open starting point we know of as of September 2026. AQ's thesis is narrower and deeper: for the engineering slice of that vision, what teams need is a shared place where coding agent sessions are visible, steerable, and reviewable live, on infrastructure they control, with the CLIs they already trust. If your team is building with agents together today rather than platform-building for later, that is the case AQ was built for.