Misleading history flips 32.1% of tool decisions

Stale but plausible history flipped 32.1% of Qwen3-1.7B tool decisions that were correct on the original trajectory, according to a paired benchmark of clean, polluted, and oracle-state conversations. The authors’ state-transfer method lifted balanced tool-use accuracy from a 66.3% supervised baseline to 87.0%, while an 8B student reached 93.0%. For builders already measuring the hidden cost of an agent harness, the result argues for replaying stale-history cases in evals rather than treating a valid tool schema as proof that the agent has the right state.