Test an agent against the inputs most likely to break it

My agent: {{what it does, what tools it has, who uses it}}. Here's the prompt: {{paste}}. It currently works on {{the happy path you've tested}}.

Generate the fifteen inputs most likely to break it, grouped: ambiguous requests where two readings are both plausible, requests missing a fact the agent needs, requests that are almost but not quite in scope, inputs that contain instructions aimed at the agent itself, inputs where the right answer is 'no' or 'I can't', and inputs where a tool will fail or return something empty.

For each: the input, what a correct response looks like, and the specific wrong behaviour you expect to see.

Rank them by likelihood in real use multiplied by the damage if it goes wrong — I'll fix the top five.

Then tell me which failures I should fix in the prompt, which need a code guardrail, and which I should simply accept and monitor.