FuguReport

AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces

Authors Sungho Park, Wonjoong Kim, Rongyuan Tan, Jue Zhang, Wook-Shin Han, Pengfei Gao, Chanyoung Park, Yongqiang Yao, Rao Fu, Elsie Nallipogu, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang
Affiliations Korea Advanced Institute of Science and Technology / Microsoft / Southern University of Science and Technology / Pohang University of Science and Technology
Categories Method / Optimization / Iterative harness updates from failures, Application / Agent Performance / Improving agent behavior, Evaluation / Benchmarking / Experimental validation on multiple benchmarks
License CC BY 4.0

Abstract Overview

This paper presents AutoSaddler, a framework for automatically improving the external harness of LLM agents on long-horizon tasks. It formulates harness optimization as an offline mini-batch learning problem over execution traces, where failures are diagnosed and converted into candidate updates. The method combines evidence-grounded diagnosis, structured patch generation over prompts, tools, and middleware, and validation-based selection with an EvoDAG history of prior updates. The goal is to produce durable harness changes that generalize beyond single trajectories rather than one-off repairs.

Novelty

The distinctive contribution is treating harness optimization as an offline learning process driven by failure traces rather than relying on manual tuning or prompt-only optimization. AutoSaddler introduces a structured intervention space spanning prompt, tool, and middleware layers with phased patch scheduling, alongside an EvoDAG-based evolution mechanism that guides candidate proposals using reflection history.

Results

Across GAIA2, SWE-Bench Pro, and Terminal-Bench 2.0, AutoSaddler improves over the corresponding base harnesses, reaching 62.0%, 46.9%, and 50.0% Pass@1, respectively, and outperforms the strongest automated baselines on each benchmark. On GAIA2, it reaches 72.3% development accuracy with approximately 1,000 task executions, whereas GEPA and Meta-Harness saturate at 64.6% and 61.5% despite consuming roughly 2,800 executions. Ablation experiments confirm that in-depth diagnosis, structured intervention, and generalization-aware selection are all essential for preventing regressions and sustaining performance gains.

Key Points

  1. AutoSaddler optimizes agent harnesses through an offline mini-batch learning loop that diagnoses failure traces and generates targeted code-level patches.
  2. The framework organizes modifications into structured prompt, tool, and middleware categories with phased scheduling, and tracks update history using an EvoDAG memory structure.
  3. Empirical evaluations demonstrate substantial gains in Pass@1 across three benchmarks along with higher optimization efficiency and lower regression rates on unseen tasks.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.