Harness of harnesses
Published August 14, 2026 · by the AQ team
A harness of harnesses is a development environment that runs multiple AI coding harnesses (the agent CLIs like Claude Code, Codex, and OpenCode) side by side under one shared layer, adding what no single harness provides for itself: isolation between sessions, persistence beyond a laptop, and a way for a team to watch, join, and review the work. A single harness manages one agent's loop; a harness of harnesses manages the harnesses.
The base term first: a coding harness is the agentic loop wrapped around a model, the system prompts, tool definitions, context management, and retry behavior that turn a raw LLM into a working coding agent. Claude Code, Codex, and OpenCode are harnesses. The term stopped being niche in 2026 when model labs started publishing benchmark results per harness; the same model can swing ten or more points across different harnesses, so the harness is now part of the measurement, not an implementation detail.
Why the term exists
Three shifts made the layer above the harness necessary. First, harness sensitivity became measurable: when DeepSeek V4 Pro reached general availability in August 2026, the loudest theme in its launch-week discussion was that scores vary sharply by harness, and Z.ai ran most of its GLM-5.3 launch evaluations inside the Claude Code harness rather than a bespoke one. Evaluating a model now means evaluating it in a harness, and picking one harness forever means inheriting its blind spots. Second, teams stopped standardizing: the best model-harness pairing changes monthly, engineers have preferences, and a team running Claude Code, Codex, and an open-weights model through OpenCode is increasingly the common case. Third, each harness ships as a single-player terminal program; everything a team needs around it (isolated checkouts, sessions that survive a closed laptop, a second set of eyes) has to come from somewhere, and it should come from one place rather than per-harness glue.
What qualifies
Running two CLIs in two terminal tabs is not a harness of harnesses. The criteria:
- Stock harnesses, not forks. The environment runs the unmodified CLIs, so upstream releases run as-is instead of waiting on a fork to catch up, and the agent behaves exactly as it does anywhere else. A fork of one agent is a single harness with a different owner.
- Real isolation per session. Each running agent gets its own checkout (typically a git worktree) so parallel sessions cannot sweep up each other's edits.
- Persistence above the harness. Sessions outlive the laptop that started them, because long agent runs are the point.
- A shared human layer. Teammates can open a running session, read the scrollback, and review the work while it happens, which is what makes session review practical.
- Model and vendor neutrality. Engineers bring their existing subscriptions, or pick models from a metered menu, without the environment taking sides.
| A coding harness | A harness of harnesses | |
|---|---|---|
| Manages | One agent's loop | Many harnesses, side by side |
| Session lives in | Your terminal | A persistent workspace a team can open |
| Isolation | Whatever directory you ran it in | One checkout per session, by construction |
| Review | Scrollback on one machine | The live session, next to the diff |
| Vendor bet | One lab's roadmap | None; harnesses are interchangeable |
How AQ instantiates it
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. Mechanically, it is a harness of harnesses: every workspace is an isolated git worktree, the stock CLIs run in persistent tmux sessions on your team's VM, teammates open the same live session from the browser once the workspace is opened to them (visibility stays owner-managed), and model spend stays on each engineer's existing Claude or OpenAI account, or on AQ Models with AQ credits priced at provider cost during early access, with $5 included at sign-up. The category name says who it is for; harness of harnesses says how it works.
Related terms
Multiplayer coding agents (the category: tools whose agent sessions are shared places), agentic development environment (the broader class of environments built for agent-driven work), agent orchestration (coordinating many agents, which a harness of harnesses gives a home), and session review (the practice the shared human layer exists to enable).
Frequently asked questions
Is a harness of harnesses just a wrapper around other tools?
No, and the distinction is load-bearing. A wrapper intercepts or modifies one agent's behavior and therefore drifts from upstream with every release. A harness of harnesses runs the stock, unmodified harnesses and builds the environment around them: isolation, persistence, collaboration, and review. The agents behave exactly as they do anywhere else; what changes is where they run and who can see them.
Why not just standardize the team on one harness?
Because the harness is now part of the result. Model labs publish benchmark numbers per harness, the same model can swing ten or more points between harnesses, and the best model-harness pairing changes as new releases land. A team locked to one harness inherits one vendor's roadmap and re-evaluates from scratch at every model release; a harness of harnesses makes trying the new pairing a menu choice.
How does a harness of harnesses relate to an agentic development environment?
An agentic development environment (ADE) is the category: an environment designed around agents doing the work. Harness of harnesses is an architecture within that category: the choice to run multiple stock harnesses under one shared layer instead of building or forking a single agent. AQ is an ADE built as a harness of harnesses.
Does a harness of harnesses replace my existing Claude or OpenAI subscription?
No, it runs on top of them. In AQ, engineers connect the agent CLIs with their own accounts and AQ never marks up that usage. Teams without subscriptions can run OpenCode or Pi on AQ Models instead: that inference consumes AQ credits, priced at provider cost during early access, with $5 in credits included when you sign up.