What are the most common AI agent failure modes?
Short answer
Agents tend to fail in a handful of recurring ways: passing wrong arguments to the right tool, reporting success on actions that actually failed, losing state across a multi-step task, taking actions they were not permitted to take, and mishandling rate limits or transient errors. Knowing these patterns tells you exactly what your tests should target.
The recurring failure modes
- Wrong arguments: the agent chooses the correct tool but passes a wrong value, so the call does the wrong thing
- False success: the agent reports a task is done when the underlying action failed, hiding the problem
- Lost state: in a multi-step task, the agent misreads what an earlier step produced and builds on a wrong assumption
- Permission blindness: the agent attempts an action it is not allowed to take and does not handle the rejection
- Fragile error handling: a rate limit or timeout derails the whole run instead of being retried
- Duplicate actions: on a retry, the agent repeats a completed action, creating duplicate records or charges
Turn the list into tests
Each failure mode maps to a scenario: a task needing a precise argument, one rejected on permissions, one that fails midway, and one that retries. Cover these deliberately and you catch the failures that otherwise reach production. The action-level failures in particular only surface when you check real state, which is why a stateful sandbox matters here.