Skip to slide
28 / 41

CH 04 · Why Multi-Agent Systems Fail: The MAST Taxonomy · 03/08

Read as one page

The contributions, in plain terms

The paper makes three contributions:

  1. MAST, a Multi-Agent System Failure Taxonomy: a careful catalog of 14 failure modes grouped into three categories.
  2. An LLM-as-a-judge method that automatically labels failures according to MAST, reaching about 94 percent accuracy and strong agreement with human experts (a Cohen's Kappa of 0.77, which is a statistical measure of agreement beyond chance).
  3. Case studies showing that you can improve real systems by fixing the failure modes MAST reveals, rather than just reaching for a bigger model.

The taxonomy itself was not invented from a whiteboard. It came from manually annotating over 200 execution traces, each averaging more than 15,000 tokens, using a careful method (grounded theory) and refining the categories until annotators agreed with each other. That grounding is what makes the taxonomy credible.

Finished this idea?

Your learning progress stays in this browser.