Can an agent recover from a tool call that returns a wrong value in the right shape?

Sometimes, if something is watching for it. A 2026 study checks each tool result against an “outcome contract,” a rule for what that value should look like, mined from the tool’s normal runs or taken straight from its schema; a negative inventory count or a cached error page returned as data both break one. When a rule breaks, nothing is blocked or rewritten. The agent gets a note naming the broken rule and the recovery tools available, and decides itself what to do next. Across four models in two model families, that lifted completion on a benchmark of injected tool failures from 10.9% to 28.1%, with a third model family replicating the result, and lifted tau-bench’s retail tasks by 12 to 14 points across DeepSeek V4 Flash and Pro. The gain only reaches as far as the contracts do: on failures outside the mined vocabulary, the agent caught less than half of them. What are the ways a single tool call can fail covers the failure itself; this is about whether anything notices once it happens.