Use Try on the Agent to check the configuration in a controlled conversation. Testing should cover ordinary success, missing information, access boundaries, and tool behavior, not only one ideal prompt.
- A common question produces a direct, useful answer
- Citations point to the expected Standard or Source
- An ambiguous question triggers clarification rather than invention
- Connector calls are proposed or executed according to policy
- A missing answer is handled honestly
Test the smallest configuration first
Start before attaching every possible Source and Area. A small scope gives clearer evidence about which instruction or attachment caused the result.
Use representative readers
Both screens run as you, so neither exercises another reader's access. An anonymous visitor is a different principal with a narrower capability set, and a widget on the real page is the only place that boundary is tested.
Inspect the evidence
Read citations and tool traces instead of judging only the prose. A fluent answer grounded in the wrong Source is still a failed test.
Retest after material changes
Changes to instructions, model, Areas, Connectors, or high-impact Sources can alter behavior. Retest the critical scenarios before publishing or updating a widget.
Next: the website widget.