Glossary

Agent evaluation

AI agent evaluation is testing whether an agent does its job correctly and stays inside its limits before it acts on real systems. It covers the quality of its answers, the tools it calls, and what happens when it should refuse.

In a run

On Ares, a limits test is a normal run with one step refused on purpose. The refusal and its reason go on the record, so the team can show the limit held.

Our word for it

Test the answers and the limits.

Two questions, two tests

Quality: does the agent give the right answer? Limits: when it is asked to do something outside its job, does it stop, and can you show that it did?

The second question needs a refusal on the record, not a report that says the agent behaved. A step refused on purpose shows the limit works before a real order is at stake.

Evaluate Ares: what you bring, and what you keep

  • One agent your team already runs, and one tool it calls.
  • A test directory, so no real user is involved.

You keep your own evidence that it works in your environment, the record of the run, and the verifier result from a check you run on your own machine. If the result holds, the next step is the licence and rollout, in the deployment shape you choose.

Where it stops

An Ares evaluation does not test whether your model gives correct answers. It covers one agent and the tools it calls, not every agent you run.

Evaluate Ares with one of your agents, or read the platform checklist guide.

Related terms

  • AI agent observabilityAI agent observability is the ability to see what an agent did, step by step, and why.
  • Refusal recordAn AI agent refusal record is the entry written when an AI agent asks to do something and is not allowed to.

Tell us about one workflow.

We will say whether an agent should do it, and what the first step costs.

A senior member of the team replies within two business days.