definition
what is an agent harness?
An agent harness is the structure around an AI agent that decides what the agent may touch, how its work is checked, and what evidence it must return. The model supplies judgment. The harness supplies tools, permissions, verification, and the record of what happened.
The term also travels as AI harness and AI agent harness. All three name the same layer: the difference between a model that answers and a system you can hold accountable for the answer.
anatomy
what a harness actually does
Strip any serious harness to its frame and four members carry the load.
tools
What the agent is allowed to call. A harness turns an open-ended model into a worker with a defined bench: named operations instead of arbitrary improvisation.
permissions
What each call may reach. Scopes, keys, and boundaries that hold even when the model is confidently wrong about what it should do next.
verification
How the work is checked before it is trusted. The weakest harnesses skip this member entirely; the strongest make it structural.
evidence
What survives the run: hashes of what was read, records of what was done, and grounds for every number that was returned.
the certifiable kind
what makes a harness certifiable
Most harnesses stop at permissions: they control what the agent touches, and take its computation on faith. A certifiable agent harness goes further on member three and four.
The common failure looks like this: an agent reaches for numpy inside a tool call, writes twelve lines over your data, and returns a number. The code was written once, gated never, and trusted anyway. Nothing visibly separates an answer that was proven from an answer that was merely produced.
A certifiable harness refuses that bargain. It does not hand-roll code on your data. It routes the data to solvers that were gated against reference implementations before they were allowed to run, cross-checks the reading between independent instruments, and returns a certificate: what was proven, by what, to what tolerance. When the instruments disagree, it names the exact hole instead of blessing the answer.
quplane is that harness for computation, shipped as an AMI on AWS Marketplace: 97 gated solvers behind one API, running entirely inside your own AWS account. It does not run your agent loop; your agent hands over a data file and a plain-language question, and the harness routes, solves, cross-checks, and hands back a verdict with the evidence still attached. See real solves with the proof or read how agents call it.
faq
questions people ask
Is an agent harness the same as an agent framework?
They overlap but pull in different directions. A framework helps you build the agent: prompts, memory, orchestration. A harness constrains and verifies the agent once it runs. Frameworks make agents capable; harnesses make them accountable.
Does a harness slow the agent down?
The gating happens before deployment, not per request: a solver is checked against its reference implementation once, then serves. At request time the harness adds routing and cross-checking, which cost seconds and buy you an answer you can act on without a human re-deriving it.
What does a certificate actually contain?
A typed contract beside the native result: status, tolerances, diagnostics, and for optimization problems an optimality bound or a completed proof. The contracts page shows the real format.
Can my existing agent use one?
If it can POST JSON, yes. Every quplane instance documents itself to agents at
/llms.txt, /agents.md, and /openapi.json;
point your agent at the instance address and it learns the API without a human in
the loop.