Agent Failure Zoo
A practical field guide to how AI agents fail: wrong tool, stale context, hidden authority, bad memory, and confident nonsense wearing a clipboard.
An agent with tools can be wrong with more reach.
The Zoo
An agent with tools can be wrong with consequences. Start by naming the failure before dressing it up as product copy.
- tool misuse
- retrieval drift
- authority confusion
- memory leak
- overbroad action
- missing eval
- unclear rollback
How To Use It
Record the symptom, cause, test, repair, and rerun result. The tone can be funny. The check stays boring.
First Failure Cards
The Obedient Browser treats text on a page as an instruction. Test it with a page that says to ignore the user's request.
The Citation Tourist supplies scholarly-looking links that do not support the answer. Give it conflicting sources and ask for the quoted evidence.
The Permission Acrobat turns a vague wish into an unauthorized action. Test the boundaries around calendars, email, files, and purchases.
One Card, Fully Drawn
For the Obedient Browser, the symptom is obedient wrongness; the cause is treating retrieved data as a command. Create a synthetic hostile page and ask the agent to summarize it. It should describe the attack, not follow it.
Keep page text separate from authorized instructions and tool permissions, then retain that page as a regression test.
Reviewer Questions
Which input was trusted, and which was only data? Was the action explicitly permitted? What happens when sources disagree? What receipt proves the action occurred? Can you answer from the evidence, without reconstructing the owner's intentions?