How do multi-agent systems actually fail?

Across three root causes, and the largest is the instruction rather than the model. The MAST taxonomy, built from more than 1,600 annotated traces across seven multi-agent frameworks, sorts 14 named failure modes into specification and system design issues (41.77 percent: the system was told to do the wrong thing, or told imprecisely), inter-agent misalignment (36.94 percent: context lost at a handoff, agents talking past each other, conflicting outputs nobody reconciled), and task verification (21.30 percent: nothing checked whether the finished work was actually correct).

None of the three is a model-accuracy problem in the sense of the model getting a fact wrong. They’re coordination and process failures: the wrong instruction going out, the wrong information crossing a seam, or no one checking the output at all. That’s also why per-agent grading misses most of it: a grader that checks each agent’s own output in isolation has nothing to say about a specification that was ambiguous from the start or a handoff nobody verified.

sources

keep reading

More on this.

Are most agent errors caused by the agent's own logic?No. In Datadog's 2026 report, 5% of LLM call spans errored in February 2026, and 60% of those were exceeded rate limits, not the agent's reasoning.answer →Can content an agent retrieves override its own instructions?Yes. Text inside a webpage, document, or tool result an agent reads can contain instructions the model follows, a failure called indirect prompt injection.answer →How do I catch an agent that claims a task is done when it isn't?Grade the state a session actually left behind, a payment reference, a ticket status, a booking, against what the user asked for, instead of trusting the agent's own closing message.answer →Everything on failure modesThe definition, and every question answered under failure modes.concept →

Send us the traces you already emit.

book a demo