Agent failure modes — taxonomy, recurring five, mitigation
taxonomies 2024-2026
MASFT (Berkeley 2025) 14 modes, 3 categories
Cohen's Kappa 0.88 on annotation reliably distinct
Microsoft Taxonomy autonomy amplifies AI failures
bias, hallucination, leakage worse when agent has tools
Characterizing Faults in Agentic AI
orchestration + state evolution + env interaction
Hallucinations Survey (arXiv:2509.18970)
Instruction deviation + Long-range context misuse
five industry-recurring modes
1. hallucinated actions fake tool, fabricated args
2. scope creep extra PRs, extra emails, extra writes
3. cascading errors one mistake, N downstream ops
4. context loss forgets early-turn constraints
5. tool misuse wrong args, wrong tool
mitigation gates at every step
per-step safety classifier (Lesson 21)
tool-call argument validation (Lesson 06)
CRITIC-style external grounding of claims (Lesson 05)
state re-probe to catch success hallucination ("was the file actually created?")
Phoenix trace clustering for drift detection in production
cascading is the killer — one phantom SKU 4 API calls multi-system incident.