Researchers from Microsoft have developed AutoSaddler, a novel system designed to automatically optimize AI agent harnesses. This method treats the harness as code and learns to patch it offline using failure traces. AutoSaddler iteratively runs tasks, diagnoses failures, generates structured patches for prompts and control logic, and validates these changes before updating. The system demonstrated significant improvements, achieving gains of 9.0 points on GAIA2, 9.6 on SWE-Bench Pro, and 10.0 on Terminal-Bench 2.0 compared to their base harnesses. AI
IMPACT Automates the optimization of AI agent harnesses, potentially leading to more efficient and capable AI systems.
RANK_REASON Paper detailing a new method for optimizing AI agent harnesses. [lever_c_demoted from research: ic=1 ai=1.0]
Read on X — Omar Sanseviero (HF research) →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →