AI agent evaluation is the practice of judging how an agent reached a result, not only what it returned. It exists because the usual method — compare the output to an expected answer — breaks down for agents: a multi-step agent can reach a good outcome by many valid routes, so there is no single correct string to compare against. Useful agent evaluation instead scores properties of the whole run: did it stay inside the rules it was given, was every claim grounded in something it actually retrieved, did it omit anything the task required, what did it cost, and did it recover when a step failed. Those properties survive the agent getting better. A fixed expected answer does not.
For a classifier or an extraction task, comparing output to a labelled answer works well. For an agent it fails in a specific way: there are many valid trajectories. Two runs can search different sources in a different order, use different tools, and both be correct. Marking one wrong because it doesn't match a stored string measures conformity and calls it quality.
It fails in the other direction too. Freeze an expected answer, improve the agent so it surfaces something the frozen answer never contained, and your suite books the improvement as a regression. The test punishes the behaviour you were trying to produce.
The dimensions that hold up across versions of an agent:
Some of those dimensions are mechanical and need no model at all — budgets, schemas, forbidden tools, loop detection. Some genuinely need language understanding, like whether a claim is supported by a retrieved passage.
The distinction worth preserving is that a model is good at finding and quoting evidence and unreliable at assigning a score. A judge asked for a number gives you a number, with no way to check it and no reason it should be the same tomorrow. A model asked to extract findings with pointers into the trace gives you something a human can audit, and lets ordinary code do the arithmetic reproducibly.
"Eval" in common usage often means a benchmark score — a number on a task set. That's useful for choosing a model. It doesn't answer whether your agent, on your data, under your constraints, is safe to put in front of customers.
Agent evaluation in the sense used here produces an artifact someone signs: a report tied to a written definition of success, with evidence for each condition, reproducible from the stored traces when the prompts or the model change.
The short answer above is ours, and we've written at length about how we got to it: