Contract2Tool: Learning Preconditions and Effects for Reliable Tool-Augmented LLM Agents

In three linesContract2Tool automatically infers tool contracts (preconditions, effects, risks, costs) from metadata, schemas, and execution traces. Learned contracts achieve 0.980 downstream success on multi-step agent tasks, near 0.990 with gold contracts, while reducing visible tools from 100 to 1 and token usage from 26,172 to 2,528.

Summary generated by Claude — human-verified