In 72% of failed multi-agent runs the log already contained evidence against the error, and in 88% of those no agent used it

The authors annotated 95 failed runs of six multi-agent systems from two public corpora, Who&When and MAD. They read each trajectory in full and recorded where the error entered, how it spread, whether anything detected it, and which control could have caught it. Faults entered early, at a median of 0.18 of the run, and in 37 of 88 Who&When runs they entered before the step the corpus's human labels mark as decisive. Something flagged a deviation in 31% of runs, but no response removed the deviated state.

By the authors' count, the controls that would have reached the most episodes were an independent judge that reads the evidence the worker did not use (74% of runs), task contracts fixed at admission (67%), re-checking beliefs against the environment (43%), stall and termination rules (40%) and provenance on stored facts (38%). Restoring the agent's own state to just before the origin would have sufficed in 35 of 95 runs. A text scan also found that 45% of the hand-crafted Who&When failures ended with a provider content-filter crash, after which the harness emitted a final answer that no agent produced.

The annotations were made by an LLM following a written protocol, and this version reports no human inter-annotator agreement. The authors say their origins fall within one step of the human-labelled decisive step in 59% of cases. The sample was chosen by annotation batch, not at random, and both corpora are question-answering and programming tasks in sandboxes, so no run changes a shared or third-party system. "Would have reached" is a judgment about the trace, not a re-run with the control in place. The protocol, records and scripts are published with the paper; we did not run them. We read the abstract, the introduction, the control results and the validity section.

Why it matters: If your own failed runs show the same pattern, a reader of the existing log and a task contract fixed at the start would be the first controls to try ahead of checkpoint and rollback. In this sample the judge and the contract reached 74% and 67% of runs, while restoring the agent's own state sufficed in 35 of 95. The caveats above apply: an LLM annotator, judged counterfactuals and sandboxed tasks.