MARS repairs failed multi-agent AI runs with tree search

MARS repairs failed multi-agent AI runs by scoring fix candidates with Monte Carlo Tree Search, per the paper posted to arXiv as 2607.29055.

That closes a gap nobody had automated: once a diagnosis tool flags which agent broke a workflow, fixing it has stayed manual work.

MARS beats prior repair methods by 3.0 to 12.1 absolute points at comparable token cost, per the paper's benchmarks.

The method narrows its search with diagnosis-guided expansion, then scores each candidate using a partial rollout instead of full simulation. That's the design choice that keeps token costs down.

The same paper releases StateMAS, a corpus of 1,310 replayable multi-agent failures spanning four agent architectures and four model backbones. That gives anyone building agent evals a standing set of failure traces to test against, not just MARS's own numbers.

I'd bet the corpus outlasts the algorithm. A tree-search method built to beat specific baselines gets leapfrogged within a year. 1,310 labeled, replayable failures across four architectures is infrastructure other methods can run against long after MARS itself is old news.

This lands as multi-agent frameworks keep shipping faster than the tools to debug them when they break.