How do you evaluate an AI agent?

I would evaluate an AI agent with a fixed suite of realistic tasks that have checkable end states. For example, a support task succeeds only if the agent opens the right ticket and attaches the correct records, not merely if its final message sounds plausible. I would define the expected state and validation rules before running the suite, then report task success rate as the primary measure. Results should be broken out by task type so a gain on easy tasks does not hide a failure on a critical workflow.

I would also measure step count, cost, latency, and error recovery. These operational metrics show whether the agent reaches the right state efficiently and whether it can recover after a failed tool call or an incorrect intermediate result. A useful test changes one condition, such as making one tool return an error, and checks whether the agent responds appropriately instead of repeating the same call. I would keep the scenario and judge consistent across versions so the comparison is meaningful.

Trajectory review is essential because final-answer scoring can hide unsafe behavior. I would inspect the tool calls, intermediate state changes, and retrieved evidence. If the agent reaches the correct answer after an unauthorized action or a risky tool call, I would count the run as a failure even though the outcome looks right. Finally, I would rerun the same suite for every model or prompt change and review regressions, rather than relying on a few manual demos.