Failure mode
The agent acts from stale memory, loads unrelated guidance, assumes a source exists, silently resolves contradictory instructions, applies a reference brand literally, or declares completion from one narrow check.
Applied correction
Define the state model before completion and test the same task under different inputs and constraints.
Field test
Did the agent verify relevant non-happy states and responsive/accessibility constraints? Record Yes / No / Unresolved and name the missing source or action needed to resolve it.
Ask your agent
Verify relevant loading, empty, partial/stale, success, validation, server error, offline/retry, disabled and permission states. Include keyboard, visible focus, semantics, accessible names, contrast, reduced motion, touch, narrow screens, long content, and 200%