AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces
AutoSaddler treats agent harness tuning as an offline learning loop driven by failure traces.
The framework diagnoses failed executions, generates structured harness patches, and keeps updates that pass validation. On GAIA2, SWE-Bench Pro, and Terminal-Bench 2.0, it improved base harness performance by 9.0, 9.6, and 10.0 percentage points. The authors say the gains depend on deeper debugging, targeted edits, and selection that favors generalization over one-off repair. HF Daily Papers' note
The framework diagnoses failed executions, generates structured harness patches, and keeps updates that pass validation. On GAIA2, SWE-Bench Pro, and Terminal-Bench 2.0, it improved base harness performance by 9.0, 9.6, and 10.0 percentage points. The authors say the gains depend on deeper debugging, targeted edits, and selection that favors generalization over one-off repair. HF Daily Papers' note
score 5