Safety

MAST (Multi-Agent System Failure Taxonomy)

MAST: per-trace annotations of 1242 multi-agent LLM system execution traces across 7 frameworks (ChatDev, MetaGPT, AppWorld, AG2, HyperAgent, Magentic, OpenManus) for 14 failure modes from the Multi-Agent Systems Failure Taxonomy. Each (system, trace) cell is labelled present/absent for each of the 14 failure modes by the LLM annotator (OpenAI o1, calibrated to kappa=0.77 with human experts).