Current multi-agent systems learn only from final outcomes, missing intermediate failures that could guide improvements. This one-shot design leaves no way to directly fix what went wrong during execution.
MAS-Loop adds a Refiner that studies the execution trajectory, the step-by-step record of how agents performed. It identifies a specific failure, generates a targeted fix, and applies it to the system. The updated system is re-run on the same task to measure the actual improvement.
The method consistently outperformed all baselines. This closed loop produces both stronger systems and a dataset of before-after pairs that can train better reward models, moving beyond trial-and-error learning towards precise, trajectory-guided self-evolution.