MAST taxonomy — the 2026 reference for multi-agent failure modes
three MAST categories (Cemri 2025, NeurIPS) — 41-86.7% failure rate on real MAS
Specification
41.77%
role ambiguity
unclear task definition
implicit success criteria
fix: explicit role contracts
Coordination
36.94%
state desync
message loss / timeout
concurrent writes
fix: versioned state + acks
Verification
21.30%
no independent checker
chain of trust
silent wrong outputs
fix: unwritable verifier
Groupthink family (arXiv:2508.05687)
monoculture
same base model → correlated hallucinations
conformity
agents align with loudest peer; rewards sycophancy
deficient ToM
fails to model each other (Lesson 18); long-horizon break
mixed-motive
partial-incentive alignment → compromise-middle outputs
cascading
retry storms; one failure amplifies through dependencies
fix: STRATUS
detection + diagnosis + validation agents → 1.5x mitigation success
+ breakers
circuit breakers on every outbound call; capped retries