Rehearse tool failures before an agent goes live

For an agent with [AVAILABLE TOOLS] doing [TASK], write six synthetic failure drills: timeout, stale browser tab, contradictory records, duplicate write response, permission denied and partial external success. For each, give setup, expected agent behaviour, prohibited shortcut, evidence to record and recovery test. Include an idempotency strategy for writes and a rule against claiming success from an unconfirmed click. Return a drill matrix and a pass/fail rubric a human can apply. Use synthetic identifiers; do not actually trigger failures in a live account or instruct the agent to bypass permissions.