Why multi-agent systems fail

A team of agents sounds more capable than one. In practice, multi-agent systems frequently underperform — studies that catalogue real systems find many fail their tasks a surprising share of the time, and often a well-built single agent does better. The reason isn’t only weak models. Coordination itself introduces failure modes that don’t exist for one agent, and naming them is the first step to fixing them.

The MAST taxonomy: three places it breaks

The Multi-Agent System Failure Taxonomy (MAST, from a 2025 study that hand-annotated hundreds of failed traces) found that failures cluster into three categories — and, strikingly, they’re spread fairly evenly across all three, so there’s no single thing to fix:

Errors cascade

On top of those categories sits a structural problem: in a chain of agents, one agent’s mistake becomes the next agent’s input. Errors don’t just occur — they propagate and compound. Even if each agent is individually reliable, stringing them in series multiplies the failure probability:

# End-to-end reliability of n agents in series, each correct with probability p.
per_agent = 0.95
for n in [1, 3, 5, 10, 20]:
    print(f"{n:>2} agents in series (each 95% reliable): end-to-end = {per_agent**n:.1%}")

 1 agents in series (each 95% reliable): end-to-end = 95.0%
 3 agents in series (each 95% reliable): end-to-end = 85.7%
 5 agents in series (each 95% reliable): end-to-end = 77.4%
10 agents in series (each 95% reliable): end-to-end = 59.9%
20 agents in series (each 95% reliable): end-to-end = 35.8%

Twenty individually-excellent (95%) agents in a row are right barely a third of the time. A single bad hand-off early can derail everything downstream — and the more agents and turns you add, the worse it gets. This is the quantitative case behind orchestration’s warning: more agents is not more reliability. (It also compounds with the conformity failure from multi-agent debate — agents can confidently agree on a wrong answer, so the “team” launders an error into consensus.)

Diagnosing and reducing failures

The fixes map onto the three MAST categories — and onto the structural cascade:

• Specification — give each agent a crisp role, an unambiguous task spec, and a strong system prompt; re-anchor the original goal so it isn’t lost across turns.
• Inter-agent — define what each agent must pass on (structured hand-offs, not raw transcripts), and make sure agents actually read their inputs.
• Verification — add explicit verification gates and sensible termination conditions so a wrong intermediate result is caught, not rubber-stamped. This is the highest-leverage fix, because it stops the cascade.
• Evaluate the system, not just the agents — trace the whole interaction (observability), and judge end-to-end task success, not per-agent plausibility.

In one breath

• Multi-agent systems often underperform single agents — coordination adds failure modes that aren’t the model’s fault.
• The MAST taxonomy sorts failures into three roughly-equal buckets: specification & design (bad roles/specs/prompts), inter-agent misalignment (poor communication, ignoring input), and verification & termination (no checking, stopping too early).
• Errors cascade: in a series of agents each mistake feeds the next, so end-to-end reliability multiplies down (20 agents at 95% each → ~36%).
• It compounds with conformity — a team can launder a wrong answer into confident consensus.
• Fixes are organization design: crisp specs, structured hand-offs, verification gates
  • termination conditions, end-to-end evaluation — and often fewer agents.

Quick check

Next

The conformity half of these failures is in multi-agent debate; the discipline of when not to go multi-agent at all is multi-agent orchestration; and measuring end-to-end success is evaluating agents.