Evals, short for evaluations, are structured tests that measure how an AI agent or model performs on a defined set of tasks, by running it on prepared inputs and scoring its outputs against expected results or criteria. They are how a team finds out whether a change made an agent better or worse before that change reaches real work.
Why do evals matter for AI agents?
Without evals, changes to an agent are judged by impression: someone tries a few questions and it seems fine. Agents change often, through new instructions, new sources, new tools and model updates, and impressions do not catch the regressions that matter.
Agents also need different evals from chat models. Answer quality is only part of it. An agent that acts must be tested on what it did: the actions it attempted, the requests it refused, the cases it escalated, whether it stayed within its role. An eval that scores only the wording of replies will pass an agent that writes well and acts badly.
How are evals run in practice?
- Offline evals: a fixed set of inputs with known good outcomes, often called a golden dataset, run before release.
- Regression evals: the same set rerun after each change, to see what moved.
- Adversarial evals: inputs designed to push the agent past its limits.
- Online evals: samples of live work scored after the fact.
Scoring can be done by rules, by people, or by another model. Model grading is fast but shares many failure modes with the model being graded, so check it against human scores before relying on it.
What does that look like in an example?
A motor insurer's claims team replaces the policy documents its agent reads. Before release, it reruns its 300-case eval. Overall answer quality is unchanged. But on the subset of cases where the agent should decline to confirm cover, it now confirms more often, because the new documents phrase an exclusion differently. The change is held, the brief is clarified, and the eval passes. Whether such a change also affects the agent's certificate is answered in does changing an agent's sources void its certificate?
How are evals different from certification?
Evals are a development tool. Builders run them, on any data that helps, as often as they like, and the results inform their next change. Certification is a decision. It is run on the customer's own filed cases, scored gate by gate against floors set in advance, and it issues a certificate that grants depth for a capability and lapses after a fixed term. Evals tell you whether a change is better. Certification tells you whether the agent may act.
Good evals make certification more likely to pass, and they can share cases. But a vendor's eval results are not a certificate for your business. Who should carry that decision is discussed in who certifies AI agents?
Which related terms should you know?
- Benchmark: a public eval used to compare models; useful for choosing, not for certifying.
- Regression: a drop in performance caused by a change.
The certification side is set out in what it means to certify an agent.
Where Pryme fits
Pryme keeps the two apart. Teams can test changes as often as they like, but authority to act comes only from Attest, which replays the agent against the customer's own filed cases at seven gates. Changes to an agent's sources or brief are recorded, so it is clear which capabilities need certifying again.
Questions people ask
Can one model grade another model's outputs?
It can, for speed, but validate it first. Compare its grades with human scores on a sample, and keep people grading anything to do with harm, refusals and escalation.
How large should an eval set be?
Large enough that each case type you care about has enough examples to show a change. For one capability, plan in hundreds rather than dozens, and give rare, costly case types deliberate extra coverage.