Definition
Plain language
A study cataloging the different ways multi-agent AI systems tend to fail.
As stated in the literature
A 2025 taxonomy paper, "Why Do Multi-Agent LLM Systems Fail?", categorizing failure modes including specification and orchestration gaps in multi-agent LLM systems.
Why it matters: Having a named taxonomy of multi-agent failures lets teams diagnose what's actually going wrong instead of vaguely calling a system 'unreliable.'
For example, MAST catalogs failures like one agent silently doing another's job, agents disagreeing on a shared goal, or the orchestrator never noticing a sub-agent has crashed.
Heard on the show
“… The authors point to prior work — a study called MAST from twenty-twenty-five — finding that this kind of failure, agents getting tangled with each other …”Episode 034 — Catching Multi-Agent Deadlocks Before Deployment With a 40-Year-Old Tool