Telling the agent to re-simulate after every edit triples re-verification and nearly triples task success
A controlled study (arXiv 2608.28147, Aug 28) ran five Qwen models over eight cases against a DWSIM simulator, three times per model-case-condition, for 120 slots per arm. The only difference between arms was one retained instruction telling the agent to request a fresh simulation after a substantive modification, with no hard gate enforcing it. Re-verification happened in 94/120 guided slots versus 32/120 unguided, cadence violations fell from 87/120 to 26/120, and bounded final success rose from 35/120 to 95/120. One model (qwen3.5-35b-a3b) ignored the instruction entirely and never succeeded, so the cheap prompt-level win is real but model-dependent. If your agent works against a simulator, test runner or any evidence source that goes stale on edit, an explicit verification-cadence line is worth more than the retry logic around it.
↳ Follow the thread