HarnessEvolve: Learning from Reference Trajectories for Reliable Agent Self-Evolution

2609.00829 ↗ · 2026-09-01 · cs.LG, cs.AI

Wen Jiang, Mingmin Chu, Yimeng Tian, Qianxin Zhang, Haofei Yang, Rui Yang, Yang Liu, Tao Lv, Fangming Li

The finding

Headline claim

A decoupled multi-agent framework (HarnessEvolve) that optimizes an agent's full harness using ground-truth-guided reference trajectories plus anti-shortcut and anti-forgetting gates beats single-component self-evolution baselines on accuracy across five benchmarks.

The test

Five benchmarks: three open-domain (SearchQA, OfficeQA, SpreadsheetBench) and two in-house Huawei enterprise datasets (CloudCoreNetwork-QA, Wireless-QA), against GEPA, ACE, and SkillOpt, using multiple models and agent frameworks.

Effect size & scope

Only headline number given: on CloudCoreNetwork-QA HarnessEvolve beats the strongest baseline by 21.6 percentage points. No per-benchmark tables or variance estimates appear in the provided text.

Limitations

Two of five benchmarks are in-house Huawei datasets not publicly verifiable; effect sizes for four of five benchmarks not reported in the text; no cost/latency figures for a multi-agent pipeline that doubles execution passes (reference trajectories); evaluation-agent reliability is assumed rather than independently measured.

Counterpoint

Corporate lab paper with proprietary datasets where the largest gain is reported; reference trajectories require ground-truth answers, so this is supervised harness tuning, not fully autonomous self-evolution; the gate agent checking for data leakage may itself leak or be gameable; gains on QA-style benchmarks may not transfer to long-horizon real-work tasks; no ablations shown in provided text.

Economy implication

Little or nothing yet for the economics of AI: this is a capability-methods paper. Its indirect relevance is that harness/scaffold quality is an optimizable layer, which slightly weakens claims that model choice alone determines agent ROI.

Who this matters for

Technical Capability: Agent builders get a concrete recipe for avoiding shortcut learning and forgetting when auto-optimizing harnesses, the two failure modes that stall self-improving agent loops.

Labor & Knowledge Work: If agent scaffolds can be reliably auto-optimized, the engineering effort to deploy and maintain agents falls, easing the staffing assumptions behind AI-ROI plans.

Abstract

Self-evolving agents advance toward autonomy by optimizing their harness---prompts, skills, tools, and execution logic---based on environmental feedback. This paradigm, however, is hampered by three challenges: \textit{credit assignment failure}, where terminal success/failure feedback makes it ambiguous which step caused the error; \textit{shortcut learning}, where agents memorize task-specific patterns rather than acquire generalizable capabilities; and \textit{catastrophic forgetting}, where unguarded updates degrade previously acquired competence. In this paper, we introduce HarnessEvolve, a self-evolving framework that learns from reference trajectories to achieve reliable agent self-evolution. HarnessEvolve decouples the execution agent from the evolutionary pipeline, assigning execution, evaluation, optimization, and gating to independent agent modules, enabling generalizable and stable harness improvements. Specifically, HarnessEvolve overcomes credit assignment failure by generating reference trajectories (execution paths produced when given the ground-truth answers) and aligning failed executions against them to extract error signals, which are clustered to reveal systematic failure patterns. To prevent shortcut learning and catastrophic forgetting, candidate harness updates must pass two gates: a quality gate that filters data leakage and prompt bloat, and a performance gate that accepts each update if it improves on the current batch without degrading recent batches, with epoch-end validation on a held-out set selecting the best-performing accepted agent snapshot. We conduct extensive experiments on several benchmarks spanning open-domain and enterprise scenarios, using different models and agent frameworks. Results demonstrate that HarnessEvolve consistently outperforms state-of-the-art baselines across all benchmarks and settings, confirming reliability across task domains.