Agents fail in recurring, recognizable ways. A short catalogue turns surprises into checklist items.
Common failure modes
- Looping — repeating a failing action without changing approach. Cap steps and detect repeats.
- Tool misuse — wrong tool or malformed arguments. Validate inputs and return useful errors.
- Context loss — dropping an earlier constraint in a long task. Restate key constraints at each stage.
- Overreach — doing more than asked. Narrow permissions and require approval for side effects.
- False completion — reporting success without verifying. Check results against the real system.
- Silent degradation — quality drops after a model or data change. Keep a standing test set and sample live output.
For each failure mode, decide the control (limit, validation, approval, monitor) and the owner before launch.