Common AI Agent Failure Modes

Agents fail differently from chatbots. The answer may look confident while the trajectory is broken.

1. Runaway loops

Symptom: repeated similar tool calls; step limit hit.
Mitigations: detect repeated actions, escalate on no state change, improve tools/errors.

2. Tool selection errors

Symptom: wrong API for the job; thrashing across tools.
Mitigations: tighter tool descriptions, fewer overlapping tools, router policies.

3. Schema / argument errors

Symptom: invalid JSON args, missing required fields.
Mitigations: strict schemas, repair loop with validation errors, typed builders.

4. Hallucinated success

Symptom: claims “email sent” when tool failed or was never called.
Mitigations: final answers must cite tool result IDs; postconditions in code.

5. Premature finalization

Symptom: answers before enough evidence.
Mitigations: required tool policies for research tasks; checklist gates.

6. Over-action / scope creep

Symptom: does extra irreversible work not requested.
Mitigations: goal contracts, allowlists, HITL for writes.

7. Prompt / tool injection

Symptom: untrusted content steers behavior.
Mitigations: isolate untrusted text, policy engine, domain allowlists.

8. Memory poisoning

Symptom: bad facts persist across sessions.
Mitigations: write policies, provenance, TTLs, user-visible memory edits.

9. Multi-agent confusion

Symptom: agents contradict, duplicate work, or deadlock.
Mitigations: clear roles, shared state schema, single writer for critical fields.

10. Eval blindness

Symptom: demos pass, production fails.
Mitigations: real-task suites, online monitoring, trajectory review.

Debugging playbook

1. Open the trajectory
2. Find first incorrect decision
3. Classify failure mode
4. Add a grader that catches it
5. Fix tool, policy, or prompt
6. Re-run suite

Summary

Most agent failures are systems failures: tools, policies, and stop conditions. Log trajectories, grade outcomes, and fix the earliest wrong branch—not only the final sentence.