Key takeaways

• Multi-agent systems often underperform a single agent using the same model, which proves the problem is coordination, not capability.
• MAST catalogs 14 failure modes in three groups: specification issues (about 42 percent), inter-agent misalignment (about 37 percent), and verification failures (about 21 percent).
• Specification issues are the largest single source of failure, meaning setup and role clarity matter enormously.
• Verifier agents often exist but check too shallowly; verification should be layered (syntax, behavior, intent).
• MAST plus an LLM judge gives a repeatable loop: collect traces, annotate, find dominant failures, intervene, re-measure.
• Fixing orchestration beats reaching for a bigger model.

Continue to Chapter 5, which reconciles the for and against arguments into a single coherent view.