Why more retries don't fix agents: diagnosis does

When a tool call fails, the common reflex is to try again. On the ToolBench-X benchmark (Tian, Shi and Zhao, arXiv 2606.25819), that reflex barely helps: extra attempts without a diagnosis added only 3.5 to 11.5 points of accuracy.

A short diagnosis hint explaining the anomaly did far better, recovering roughly 60 to 80 percent of the lost accuracy across five models.

These are benchmark results, not Keeprun results. But they are the reason Keeprun returns a diagnosis note rather than encouraging a retry loop.