A team of AI agents can finish a coding job, announce it's done, and hand you a program that breaks the first time you run it.
Before you blame the model, look at how the agents decided they were finished.
A multi-agent system splits one task across several AI agents (language models given roles, like planner, coder or reviewer) that message each other. A UC Berkeley team collected over 1,600 annotated traces (the full log of what every agent said and did during one task) from 7 popular frameworks, and sorted what broke into 14 named failure modes, a classification called MAST, in three buckets:
1) System design: how roles and rules were set up. The most common failure lives here: step repetition, an agent redoing work it already did (15.7% of labeled failures). Not knowing when to stop: 12.4%.
2) Inter-agent misalignment: the agents lose the thread with each other, or with their own plan. The big one is reasoning-action mismatch (13.2%): the agent's stated plan says one thing and its next action does another.
3) Task verification: nobody checks the result properly, or the check itself is wrong.
Stopping and checking are where a lot of this goes wrong, so look there before you reach for a bigger model. In the paper, many verifier agents (the ones whose job is checking) ran only superficial checks, like whether the code compiled (turned into a program without errors). One chess program passed those checks and still had bugs once it ran, because nothing tested it against the rules of chess.
A reviewer agent that only reads the worker's report can repeat the worker's mistake, since both see the same claims. A check that runs something real, like automated tests, gives a signal no agent can talk its way past. I haven't seen a clean measurement of how much that beats a second AI reviewer, so test it on your own traces.
Quick check before you scroll: An agent's plan says "run the tests" but its next action edits a file instead. Which MAST failure mode is that, and which of the three buckets is it in?
If this changed how you'd debug a multi-agent setup, a repost is the only way it reaches people who don't follow me yet, and plenty of them are building these systems right now.
Full breakdown + the answer: frankduah.me/learnings/2026-10-19-multi-agent-failure-modes
New here? I post a bite-size AI / ML concept like this every day. Follow me for the daily drop, and it compounds fast. Why I do it: https://lnkd.in/gK8knHDH
#MultiAgentSystems #AgenticAI #LLMEvaluation #AI #LLM #AIAgents #MachineLearning
The answer
Reasoning-action mismatch, which sits in the inter-agent misalignment bucket. It was the second most common mode in the paper at 13.2%.