When a team of AI agents gets a task wrong, the instinct is to blame the model: it should have been smarter. A 2025 paper from UC Berkeley and collaborators, Why Do Multi-Agent LLM Systems Fail?, looked at the question systematically, and the answer is mostly something else.

The study

The authors collected traces from seven popular multi-agent frameworks running coding, math and general agent tasks. They read 150 of them closely to build a taxonomy of failures, checked that different annotators applied it the same way (a Cohen's kappa of 0.88, which is high agreement), and then used it to label more than 1,600 traces.

They called the result MAST, the Multi-Agent System Failure Taxonomy. It has 14 failure modes in three groups:

System design issues. The agents were set up badly: unclear roles, a task specification they did not follow, steps repeated, no sense of when to stop.
Inter-agent misalignment. The agents were set up fine but did not work together: one withheld information another needed, ignored another's input, drifted off the task, or carried on without asking for clarification.
Task verification. Nobody checked the work properly: the system stopped too early, verified incompletely, or verified the wrong thing.

The striking part is how little of this is about raw intelligence. Most of these failures would happen to a team of very capable people too, if they had vague roles, no shared notes and nobody reviewing the result.

A practitioner's view

Anthropic described the same pattern from the inside when it wrote up its multi-agent research system in June 2025. Early versions had subagents that "misinterpreted the task or performed the exact same searches as other agents." The lesson they drew: "Without detailed task descriptions, agents duplicate work, leave gaps, or fail to find necessary information."

The same write-up shows why the effort is worth it. On their internal research evaluation, the multi-agent system outperformed a single agent by 90.2%. It also used about 15 times more tokens than a chat. Coordination is where both the gains and the waste come from.

What to do about each kind of failure

The three groups map onto three habits you can build into any multi-agent setup.

For design failures: write the rules down, once, where agents read them. Each agent should know its role, what it may and may not do, and what done looks like. Put this in versioned instructions rather than in a prompt someone pasted once. When the rules change, every agent should be on the new version, and you should be able to tell which ones are not.

For misalignment: give agents a shared record instead of private channels. Withheld information and ignored input are what happens when each agent only sees its own conversation. If the decisions, the findings and the open questions are in one place every agent reads, there is far less to withhold and much less to repeat. Addressing a specific agent, instead of broadcasting to all of them, keeps that record readable.

For verification failures: keep a person in the loop, at the right moments. The paper's third group is about stopping early and checking poorly. Agents are getting better at checking each other, but the cheapest verifier is still a human who can see the work as it happens, and who is asked only when a decision actually needs them.

How this shapes nubi

These three habits are the design of nubi. Skills and skill sets are the written rules, versioned, and attached to an agent as its mandate; nubi can tell when an agent is following an old one. Rooms are the shared record, with mentions so agents are addressed rather than flooded. And you are a member of the room, so verification is a matter of reading along, not of asking each agent what it did.

None of that makes the model smarter. The research suggests that, for teams of agents, that is usually not what is missing.

Sources

Mert Cemri et al., "Why Do Multi-Agent LLM Systems Fail?" (2025). arXiv:2503.13657
Anthropic, "How we built our multi-agent research system" (June 2025). anthropic.com