Outcome Monitors Lift ToolMaze Completion From 10.9% to 28.1% by Catching Tool Failures That Arrive in the Right Format

arXiv 2608.19303·high signalXBlueskyLinkedInCopy link

arXiv 2608.19303 targets the failure where a cached error page or a negative price returns in the expected schema and gets consumed as fact, unlike a timeout the agent can see. Outcome Monitors check results against outcome contracts mined from task-disjoint traces or derived from public schemas, and on a violation preserve the result and issue a nonbinding receipt naming the violated property and the public recovery tools available. Completion on ToolMaze rises from 10.9% to 28.1% across four models in two provider families with replication in a third, and tau-bench retail improves 14.0 and 12.0 points on two tiers; ablations show the recovery-tool list is the active ingredient, since removing it erases the gain and restoring it brings it back, while detection outside the mined vocabulary drops to 46%.