Which MAST failure modes show up most in production?

Step repetition is the single most common of MAST’s 14 named modes, a step that already ran, running again, and it shows up in about 16% of the paper’s more than 1,600 annotated traces. Reasoning-action mismatch, where an agent’s stated reasoning doesn’t match the action it actually takes, is close behind at about 13%, and an agent that never checks its own termination condition accounts for about 12%. The three top-level categories put system design issues ahead overall, and these two modes, both system design issues, are most of why.

The ranking isn’t a fixed law. The paper’s own per-framework breakdown shows AppWorld dominated by premature termination instead, likely tied to its star topology with no predefined workflow, and OpenManus dominated by step repetition specifically. Read the headline ranking as where to look first in an unfamiliar trace, not as a rule your own multi-agent system has to follow.