How to evaluate an AI agent before customers depend on it.

A traditional chatbot can often be judged from its answer. An agent may search, call tools, update systems, recover from errors, and ask a person for approval before it produces a final message. Evaluating only that final message misses most of the system.

A useful evaluation examines the complete trajectory: which tools were selected, what arguments were used, whether permissions were respected, how the system handled unexpected results, and whether the business ended in the correct state.

Start with successful examples, then add the cases that make experienced employees slow down: missing data, conflicting instructions, duplicate records, policy exceptions, unavailable tools, and requests that require escalation. Synthetic scenarios are useful, but production logs and team interviews reveal the failures that actually matter.

Each scenario needs an observable definition of success. ‘Helpful’ is too vague. ‘Creates the ticket with the correct priority, includes the account ID, and does not expose private notes’ can be tested.

One aggregate score can hide a dangerous weakness. Separate business outcome, reasoning trajectory, tool behavior, communication quality, safety, latency, and cost. That lets the team see whether an apparent improvement merely moved the failure elsewhere.

Models, prompts, tools, and business policies all change. A production evaluation set should run when any of them changes, and live failures should become new regression cases. Trace review helps distinguish a model problem from a bad tool contract or missing context.

The goal is not to prove that an agent is perfect. It is to know where it is dependable, where it needs oversight, and whether each release improves the outcomes that matter to the business.