Agent evaluation (evals)
Qualityevals · AI evaluation · agent test harness
A repeatable set of cases with expected outcomes, run against an agent before and after every change. It measures whether the agent does the client's job, not whether it sounds good. It is what turns "seems to work" into a number you can defend.
An agent evaluation combines a dataset of representative cases, an explicit success criterion and an automatable run. It can be deterministic (does the reconciled amount match?), human-reference based, or model-as-judge. Stanford's AI Index 2026 documents the jump in public agentic benchmarks — OSWorld, SWE-bench Verified — and also the gap between benchmark performance and production outcomes. Implementa's editorial position: public benchmarks are for picking a model, not for accepting a delivery. The eval that counts uses the client's real cases and the client's definition of correct. Running evals before and after each deploy is what lets you catch quality drift instead of hearing about it in a complaint.
How it differs from
- Public benchmark
- Compares models against each other; says nothing about the client's case.
- Manual QA
- Not repeatable or comparable across versions; an eval is.
Examples
- 200 historical tickets with the correct resolution annotated, run on every release of the support agent.
- A set of 50 contracts with the clauses the agent must extract, with zero tolerance for false negatives.
FAQ
- How many cases do I need?
- Enough to cover the failure modes that hurt. Start small and grow the set with every real incident.
- Can a model be the judge?
- Useful for subjective criteria, but calibrate it against human judgement before trusting it.
Related terms
Sources & further reading
Stanford HAI — AI Index Report 2026, Chapter 2: Technical Performance · 2026-04-13 · E3