Agent Harness开始自动修复：系统级Debug最高提升18.4点

TL;DR - HarnessFix diagnoses and repairs system-level flaws in LLM agent harnesses by linking failed execution traces to editable components. Across four benchmarks, it improved task completion rates by 6.3–18.4 percentage points over initial harnesses.

• HTIR reconstructs data/control flow and maps failures to prompts, tool schemas, controllers, logging hooks, or validators.
• Diagnoses recurring flaws across execution, tools, memory, lifecycle, observability, verification, and governance.
• Scoped repair operators constrain patches and require validation for targeted improvements and regressions.
• Repairs developed with GPT-5 mini transferred across four other model families, yielding gains of 5.5–9.5 points on GAIA.