What is the MAST taxonomy?

MAST is a taxonomy of 14 recurring ways multi-agent LLM systems fail, sorted into three categories: system design issues, inter-agent misalignment, and task verification. It comes from a Berkeley paper that built it by annotating traces from seven open-source multi-agent frameworks and clustering the failures its human annotators agreed on, with a Cohen’s kappa of 0.88.

Most of the 14 modes live between agents rather than inside any single agent’s output: a step gets repeated, a handoff drops context, an agent stops before the work is done, or a result gets accepted with no check. Grading each agent’s final answer on its own misses all of that, which is why MAST classifies traces instead of outputs.

To use it on a production failure, walk the trace step by step and check what each handoff actually passed and what verified the result before it shipped. The category you land in points at the fix: a design issue means a prompt or role change, a misalignment means changing what agents pass each other, a verification failure means a missing check. How multi-agent systems actually fail breaks down which category shows up most in practice.