MAST Multi-Agent Failure Taxonomy
Annotated traces sort multi-agent failures into system design, inter-agent misalignment and task verification, and targeted prompt or topology fixes give only modest gains.
Evidence: strong
When a system of cooperating agents fails, it is easy to blame the model. The MAST (Multi-Agent System failure Taxonomy) study asked instead where such failures actually come from, by annotating real traces.
The Berkeley-led team annotated 1,642 traces across seven multi-agent frameworks and arrived at fourteen failure modes in three categories:
| Category | Share of failures |
|---|---|
| System design issues | 44.2% |
| Inter-agent misalignment | 32.3% |
| Task verification | 23.5% |
Annotation was reliable, with human inter-annotator agreement of κ = 0.88 and an LLM annotator reaching κ = 0.77. The three categories partition the taxonomy itself, so the shares show how failures distribute across those categories. They do not compare system-caused failures against model-caused ones. "Design matters more than the model" is the authors' interpretation, not a measured comparison.
The authors also tried two targeted interventions on one framework (ChatDev) with the underlying model fixed. Improved role specification in the prompts gave +9.4 points, and a changed topology with stronger verification gave +15.6 points. Both changed the prompts or the topology, so neither is a controlled test of a single factor. The authors describe the gains as modest and read them as a sign that deeper redesign is needed.
This is a peer-reviewed benchmark paper with good agreement figures, so the category split is reasonably trustworthy for the frameworks it covered.
Read as a practitioner's checklist, the taxonomy points at the design and verification layers, which you control. Where several agents act on the same systems, coordination is itself a source of failure. That argues for single-threaded writes, for recording every cross-role decision on a shared task board, and for explicit verification before done. MAST's categories give a vocabulary for incidents involving more than one agent and a seed for your own error analysis. The step-level view is in root-cause step localisation, and runtime fault classes are in the agent runtime fault taxonomy.
Sources
- Cemri, Pan, Yang et al., arXiv 2503.13657, NeurIPS 2025 Datasets and Benchmarks (MAST) — https://arxiv.org/abs/2503.13657