Verify the protocol boundary and the user experience together.
1. Contract tests
Use a fake server to verify initialization, discovery and tool responses. Include schema changes and unsupported capabilities rather than only a happy path.
2. Agent evaluation
Score whether the correct authorized tool was selected and whether the answer follows the returned evidence. Fluent answers can still cite the wrong source.
3. Operational fallback
Handle server outages with cached approved information or a clear unavailable state. Do not route private requests to an unapproved replacement server.
Worked scenario
A server outage leaves the app’s manual document browser usable. The assistant explains that live lookup is unavailable.
Apply it
Run a release matrix covering network loss, authorization failure, malicious content and correct citations.
Check your understanding
You can demonstrate a useful fallback without weakening the trust boundary. Explain the decision and show evidence from your implementation or design. If you cannot demonstrate it yet, revisit the relevant section before continuing.