Protocol and schema
Catches changes that break a connected agent. We check that client and server agree on a protocol version (version negotiation), flag tools added to or removed from the catalog, compare schemas with the accepted release, and validate structured outputs.
Tool decisions
Catches an agent using the wrong tool. A fixed set of test prompts (golden prompts) checks that the agent picks the right tool, avoids tools it should not use, passes correct arguments, and gets the expected result.
Permissions
Catches actions that should not go through. We check that tokens were issued for this interface (audience) and that sign-in and scopes are enforced. Writes must wait for confirmation, a repeated request must not apply a change twice (replay and idempotency), and actions based on outdated data must be refused (stale state).
Performance and recovery
Catches slowdowns and weak recovery. We track latency, error rate, and recovery after failures, keep a record of each test run, and compare the results with the accepted release.