# Multi-Task Self-Evolving Harnesses with the Agent as Its Own Optimizer

DevFeed: [Multi-Task Self-Evolving Harnesses with the Agent as Its Own Optimizer](<https://devfeed.tech/articles/multi-task-self-evolving-harnesses-with-the-agent-as-its-own-optimizer-131977.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.38372>)

Author: Qiankai Xu

Published: 2026-11-13T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Agent Harness](<https://devfeed.tech/topics/agent-harness.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [AI Context Management](<https://devfeed.tech/topics/ai-context-management.md>)

Tags: [agent](<https://devfeed.tech/tags/agent.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [claude-code](<https://devfeed.tech/tags/claude-code.md>), [codex](<https://devfeed.tech/tags/codex.md>), [compaction](<https://devfeed.tech/tags/compaction.md>), [completion](<https://devfeed.tech/tags/completion.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [distribution](<https://devfeed.tech/tags/distribution.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [evolution](<https://devfeed.tech/tags/evolution.md>), [framework](<https://devfeed.tech/tags/framework.md>), [glm](<https://devfeed.tech/tags/glm.md>), [glm-5-3](<https://devfeed.tech/tags/glm-5-3.md>), [glm-5-3-flash](<https://devfeed.tech/tags/glm-5-3-flash.md>), [gpt](<https://devfeed.tech/tags/gpt.md>), [gpt-5-6](<https://devfeed.tech/tags/gpt-5-6.md>), [gpt-5-6-sol](<https://devfeed.tech/tags/gpt-5-6-sol.md>), [harness](<https://devfeed.tech/tags/harness.md>), [pretraining](<https://devfeed.tech/tags/pretraining.md>), [recursive-self-improvement](<https://devfeed.tech/tags/recursive-self-improvement.md>), [training](<https://devfeed.tech/tags/training.md>)

## AI overview

The framework has one frozen model solve tasks and then directly edit the harness using complete run records. It evolves a shared harness across tasks from five benchmarks, separates training from held-out tasks, and tests on five additional out-of-distribution benchmarks. Evolution proceeds in two stages: multi-task pretraining and continual training. From a 49-line seed, GLM-5.3 and GPT-5.6 Sol harnesses gain about five points on in-distribution benchmarks, matching or exceeding Claude Code and Codex; GPT-5.6 Sol is on par with Codex on the five out-of-distribution benchmarks. Continued evolution on Claw-Eval lifts GPT-5.6 Sol and GLM-5.3-Flash above those systems. The abstract identifies output truncation, history compaction, completion checks, independent review, and visual inspection as mechanisms that emerged. These are reported abstract results; detailed methods and evidence are not included in the supplied text.

## Source excerpt

arXiv:2609.38372v2 Announce Type: replace Abstract: A harness is the code around a language-model agent that organizes prompts, calls tools, manages context, and controls execution. As models grow stronger, recent work has begun to let agents improve their own harnesses, a line of work known as self-evolving harnesses. In most existing methods, a separate proposer running on a human-designed harness modifies the solver's harness, and a separate harness is evolved for each benchmark. Real-world tasks come from many domains, so both the evolution and the evaluation of a harness should cover a diverse range of tasks. We propose a framework close to recursive self-improvement: the same frozen model, on the same version of the harness, first solves tasks as the solver and then, as the proposer, reads the complete run records and directly edits the harness that runs it. Each evolution batch draws tasks from five benchmarks in different domains. To measure generalization, training and held-out tasks are strictly separated, and we additionally evaluate on five out-of-distribution benchmarks never used during evolution. We frame the evolution process as deep-learning training with two stages, multi-task pretraining and continual training. Starting from a 49-line seed harness, the harnesses evolved by GLM-5.3 and GPT-5.6 Sol gain about 5 points on the in-distribution benchmarks, matching or exceeding Claude Code and Codex, and the GPT-5.6 Sol harness is on par with Codex on five out-of-distribution benchmarks. Continued evolution on Claw-Eval in the second stage then lifts GPT-5.6 Sol and GLM-5.3-Flash above Codex and Claude Code. We also provide an in-depth analysis of the mechanisms that emerged during evolution, including output truncation, history compaction, completion checks, independent review, and visual inspection. Code and the evolved harnesses are available at https://github.com/xqkkk/self-evolving-harness.