AI agents often trust bad data if it looks right, and a new paper shows a simple fix cuts that blind spot dramatically.
Researchers introduce Outcome Monitors, a system that checks whether a tool call's result actually satisfies expected properties, rather than just checking whether the call returned without error. When a violation shows up, like a cached error page or a negative price slipping through as valid data, the monitor flags it and points the agent toward recovery tools instead of silently passing along bad output. In testing across four models from two provider families, this pushed completion rates on the ToolMaze benchmark from 10.9% to 28.1%, more than doubling them, and the effect held up in a third provider family. On the tau-bench retail benchmark, completion improved by 14 and 12 points across two difficulty tiers.
The interesting part is what the researchers ruled out. Stripping the list of recovery tools from the monitor's output erased the gains entirely, and adding it back restored them. Diagnostic detail and timing information made no measurable difference. That suggests the benefit comes specifically from telling the agent what to do next, not from telling it more about what went wrong.
It's a useful reminder that today's agents are largely bad at distinguishing plausible-looking data from correct data, and that fixing this may have less to do with better error messages than better next steps. The catch: detection accuracy dropped to 46% on failure types outside the system's known vocabulary, so this works well against failures researchers anticipated, not the ones nobody wrote a rule for yet.