AutoSaddler Automates Agent Harness Repair, Boosting SWE-Bench by 9.6 Points

omarsar0 · x · 2026-08-25

Microsoft introduces AutoSaddler, a method to automatically optimize agent harnesses. It treats the harness as code, learning to patch it offline from failure traces by generating structured updates to prompts, tool configs, and control logic.

The system runs mini-batches of tasks, diagnoses failures, and validates updates. It achieved gains of 9.0 points on GAIA2, 9.6 on SWE-Bench Pro, and 10.0 on Terminal-Bench 2.0. The paper finds that deep debugging beats shallow reflection and targeted edits outperform unconstrained ones.

Original post →

More from coding & agent

coding & agent channel →