Testing agents
Preview a draft in chat, run its test scenarios and see exactly what each run did and cost.
Preview chat
Start a chat in Preview mode to talk to the effective draft, before it's published. The chat shows the published and deployed versions separately, the model and the budget per run.
Up to four scenario buttons from the agent's own test suites fill the message box for you. Developer mode shows the metadata of each run: the model, tokens, cost and tools called.
Test suites
Every agent has test cases: a message and what the answer must or must not do. Run the suite from the agent's Tests page; results show which passed and why others failed. A test run of the same agent can't run twice at once, and runs per hour are limited.
Official agents come with suites for core behaviour, knowledge, tools, handoff, safety and regressions. A failing required test blocks publishing.