Microsoft researchers present AutoSaddler, a framework that treats agent-harness improvement as an offline learning problem using failure-trace diagnosis, structured patch generation that treats the harness as code, and validation-based update selection. The method iteratively refines harnesses through mini-batch failure signals rather than manual tuning. Evaluation across GAIA2, SWE-Bench Pro, and Terminal-Bench 2.0 shows performance improvements of 9.0, 9.6, and 10.0 percentage points respectively.
