Notes from building agent reliability.

On ToolBench-X, a short diagnosis hint recovered roughly 60 to 80 percent of lost accuracy; extra attempts without a diagnosis added only 3.5 to 11.5 points.

Per-run success looks fine until every run has to succeed. What that means for agents that act.

What we saw running Keeprun on Platestory's agents. Real numbers only, once we have them.