agent evaluation
Testing what an agent can actually do, safely, before trusting it with real work.
An agent evaluation gives an agent realistic tasks and measures how often it succeeds, fails or misbehaves. Tests run in a sealed environment so mistakes cannot reach real systems. The seal itself must be tested, because a capable agent may find a way out. Results show where an agent is ready and where it needs a person watching.
Think of it as
Like a driving test on a closed course before the learner is allowed onto the motorway.
Example
Before launch, a company runs its refund agent on five hundred past cases in a test copy of its systems. It gets most right but over-refunds on damaged items, so that case is sent to a person.