Failure Modes — MAST, Groupthink, Monoculture, Cascading Errors

> The reference taxonomy for 2026 is MAST (Cemri et al., NeurIPS 2025, arXiv:2503.13657), derived from 1642 execution traces across 7 state-of-the-art open-source MAS showing 41–86.7% failure rate. Three root categories: Specification Problems (41.77%) — role ambiguity, unclear task definitions; Coordination Failures (36.94%) — communication breakdowns, state desync; Verification Gaps (21.30%) — missing validation, absent quality checks. The Groupthink family (arXiv:2508.05687) adds: monoculture collapse (same base model → correlated failures), conformity bias (agents reinforce each other's errors), deficient theory of mind, mixed-motive dynamics, cascading reliability failures. Cascading example: retry storms where a payment failure triggers order retries, which trigger inventory retries, which overwhelm inventory service (10x load in seconds — needs circuit breakers). Memory poisoning: one agent's hallucination enters shared memory, downstream agents treat it as fact; accuracy decays gradually, making root-cause diagnosis painful. STRATUS (NeurIPS 2025) reports 1.5x mitigation-success improvement via specialized detection / diagnosis / validation agents. This lesson treats failure modes as first-class engineering targets.

Type: Learn

Languages: Python (stdlib)

Prerequisites: Phase 16 · 13 (Shared Memory), Phase 16 · 14 (Consensus and BFT), Phase 16 · 15 (Voting and Debate Topology)

Time: ~75 minutes

Problem

Multi-agent systems fail 41-86.7% of the time on real tasks (Cemri et al. 2025 measured this across 7 open-source MAS). That is not debuggable by "just add more agents." The failures have structural causes. The MAST taxonomy gives you the categories. This lesson maps each category to a concrete detection, diagnosis, and mitigation pattern so the numbers stop looking arbitrary.

The 2026 production practice is to treat failure modes as design inputs. Your architecture is not "good enough" until you can point to each MAST category and name the mitigation you deployed.

Concept

MAST categories

Specification Problems (41.77% of failures). The agent's task was not defined tightly enough. Examples:

• Role ambiguity: two agents both think they are the reviewer.
• Task underspecified: "summarize this" when the user wanted a specific angle.
• Success criteria implicit: the agent cannot tell if it succeeded.

Mitigations:

• Write explicit role contracts. Each agent's prompt states what it does *and what it does not do*.
• Acceptance tests per task. Before the agent starts, define "done looks like X."
• Pre-flight spec check: a separate agent reviews the task definition before dispatch.

Coordination Failures (36.94%). Communication or state breakdowns.

Examples:

• Two agents update shared state without synchronization.
• Message lost between agents (queue failure, timeout).
• State drift: agent A thinks the task is done; agent B is still executing.

Mitigations:

• Versioned shared state with optimistic concurrency.
• Explicit acknowledgment for critical messages (retry until acked).
• Periodic state-sync checkpoints; detect drift early.

Verification Gaps (21.30%). No independent check on outputs.

Examples:

• One agent claims success; no one verifies.
• Chain of agents each trusts the prior's output.
• Test coverage missing on the emergent composed behavior.

Mitigations:

• Independent verifier agent (Lesson 13). Read-only, independent source access.
• Explicit handoff contract: "A's output must pass checker C before B starts."
• Outcome logging for post-hoc analysis.

Groupthink family (arXiv:2508.05687)

Five related failures when agents homogenize or mimic each other:

Monoculture collapse. Same base model or training data → correlated errors. When three agents share an LLM, they share its hallucinations.

Conformity bias. Agents adjust toward the loudest or most-confident peer, even when wrong.

Deficient ToM. Agents fail to model each other's beliefs; coordination falls apart (Lesson 18).

Mixed-motive dynamics. Agents with partially-aligned incentives drift toward compromise-middle, which satisfies no one.

Cascading reliability failures. One component's error pattern triggers error patterns in dependent components.

Cascading example — the retry storm

A classic 2026 incident pattern:

payment service fails 10% of requests
   ↓
order agent retries payment (exponential backoff but naive)
   ↓
each retry is a new order-inventory check
   ↓
inventory service sees 2x normal load
   ↓
inventory service starts timing out
   ↓
every order retries inventory check
   ↓
inventory service sees 10x normal load
   ↓
cluster goes down

The fix is classical: circuit breakers. When downstream error rate exceeds threshold, short-circuit with cached or default results. Plus capped retry budgets per request.

Circuit breakers are one of the few multi-agent failure mitigations you borrow directly from distributed systems without modification.

Memory poisoning (revisited)

From Lesson 13: one agent's hallucination becomes shared-memory fact; downstream agents reason on the poisoned fact. In MAST terms, this is a verification gap at the shared-memory layer.

Gradual accuracy decay is the symptom. You do not get a crash; you get slow drift that is hard to root-cause.

Mitigation: append-only log, provenance, unwritable verifier. Already covered in Lesson 13.

STRATUS — specialized agents for failure detection

STRATUS (NeurIPS 2025) reports 1.5x mitigation-success improvement when you deploy:

• Detection agent. Watches for symptom patterns (high disagreement, retry spikes, accuracy drift).
• Diagnosis agent. Given symptoms, infers likely root cause from the MAST taxonomy.
• Validation agent. After a mitigation is applied, checks that symptoms clear.

This is SRE-style incident response, applied to agent systems. The three roles can all be LLM agents with specialized prompts.

The failure-mode audit

A 2026 best practice is an annual (or per-major-release) failure-mode audit:

1. Trace sample. Collect ~1000 real execution traces.
2. Categorize. For each trace's failures, map to MAST + Groupthink categories.
3. Compute failure-by-category rate. Which categories dominate your system?
4. Rank mitigations. Which fix would eliminate the most failures?
5. Pick 2-3 mitigations. Implement; re-audit next quarter.

The discipline is more important than the specific choices. Without audits, failures blend into noise and never get systematically addressed.

When systems fail silently

The most dangerous failure category is silent correctness failure. A system that fails loudly (crash, exception, alert) can be monitored. A system that produces plausible-but-wrong outputs cannot be detected by exception logs. This is why verification gaps are the most expensive category per-failure even though they are only 21.30% by count.

Invest in:

• Sample-based human review.
• Golden-dataset regression tests.
• Cross-agent cross-checking on important outputs.

Failure vs slow failure

Some failures are immediate; some are slow. Immediate failures (timeout, schema mismatch, auth error) are cheap to detect. Slow failures (memory poisoning, monoculture drift, role ambiguity) are expensive to detect and prevent.

The 2026 engineering move: instrument slow-failure proxies so you can catch drift before it becomes a visible error. Agreement rate, retry rate, output-length distribution, and edit-distance between consecutive agent versions are all useful proxies.

Build It

code/main.py implements:

• FailureTaxonomy — categorizes simulated incidents into MAST + Groupthink categories.
• CircuitBreaker — classic pattern; opens when error rate exceeds threshold.
• RetryStormSimulator — shows the cascading failure; toggles circuit breaker on / off.
• DetectionAgent — scripted STRATUS-style symptom matcher.

Run:

python3 code/main.py

Expected output:

• retry storm with no circuit breaker: inventory errors blow up (simulated).
• with circuit breaker: cap at threshold; degraded-mode responses served.
• detection agent flags the pattern and names the MAST category.

Use It

outputs/skill-mast-auditor.md runs a MAST-style failure-mode audit on a multi-agent system. Traces → categorization → mitigation ranking.

Ship It

Failure-mode discipline in production:

• MAST audit per quarter. Not annual. Categories shift as your system grows.
• Circuit breakers everywhere. Each outbound call to any dependent service. Default open threshold at 5-10% error rate.
• Golden datasets. Small, high-quality, hand-audited. Regression-test against them weekly.
• STRATUS trio. Detection + Diagnosis + Validation agents monitoring production. Start with the detection agent only; add diagnosis when symptoms are noisy.
• Failure budget. Explicit SLO for failure rate by category. Exceeding budget triggers a stop-shipping conversation.

Exercises

1. Run code/main.py. Confirm the circuit breaker caps the retry storm. Vary the failure threshold and observe the tradeoff.
2. Implement a slow-failure proxy: agreement rate across 3 parallel agents. When it drops sharply, trigger an alert. Simulate a monoculture drift by gradually correlating agent outputs.
3. Read Cemri et al. (arXiv:2503.13657). Pick one of their 7 MAS systems and map its top 3 failure categories. How do these compare to what MAST predicts?
4. Read the Groupthink paper (arXiv:2508.05687). Identify which of the five patterns is hardest to detect in production. Propose a proxy metric.
5. Design a STRATUS-style detection-diagnosis-validation trio for a specific multi-agent system you know. Which symptoms does detection watch for? What mitigations does diagnosis recommend? How does validation confirm they work?

Key Terms

Further Reading

• Cemri et al. — Why Do Multi-Agent LLM Systems Fail? — MAST taxonomy, NeurIPS 2025
• Groupthink failures in multi-agent LLMs — monoculture, conformity, and the five-family taxonomy
• STRATUS — specialized agents for MAS incident response — NeurIPS 2025 proceedings entry (detection + diagnosis + validation)
• Release It! — stability patterns (Nygard) — the canonical circuit-breaker reference
• Anthropic — Multi-agent research system — production failure-mode notes