Why do multi-agent systems fail, and how do you make them reliable?

In detail

Failure lives between the agents

Each agent in a system can work in isolation and the system can still fail, because the weak points are the connections. An agent hands off with missing context. One agent’s hallucination becomes the next agent’s input, which it treats as fact. A small error early in a chain gets amplified at every downstream step. The failure is emergent — it belongs to the system, not to any one component.

Researchers at UC Berkeley put numbers on this in Why Do Multi-Agent LLM Systems Fail?. They start from the observation that multi-agent systems often post only minimal gains over single agents on popular benchmarks, then annotate more than 1,600 execution traces from seven popular frameworks. The resulting taxonomy has 14 failure modes in three groups: problems in how the system is designed, misalignment between agents, and weak verification of the task’s result.

Cascading and systemic risk

Because outputs feed inputs, multi-agent systems turn small faults into large ones. A single mislabeled result, a tool that returns stale data, a compromised agent — any of these can propagate through the whole network. The same property makes them a security surface: an attacker who influences one agent can steer the others. Reliability and security are the same problem viewed from two sides.

Cost compounds too

The gaps cost money as well as accuracy. Anthropic reported that in its own data, agents use about 4 times as many tokens as chat interactions and multi-agent systems about 15 times as many, so they pay off only on tasks valuable enough to cover that. The same post notes that work where every agent needs the same context, or where agents depend heavily on each other, fits multiple agents poorly. A loop between agents that never converges, like the one Lyzr stopped at $300, is the expensive version of that mismatch.

Making them reliable

Treat the system as the unit of analysis. Trace every step so you can see where a failure entered and how it spread; what is AI observability covers what to record. Validate what passes between agents instead of trusting it. Put guardrails on how much autonomy each agent has. And threat-model the cascade paths in advance — ask how a fault in agent A reaches the user — so you catch systemic failures in design rather than in production.