Evo$^2$: Evolving the Evolver through Weight-Harness Co-Evolution

Chejian Xu; Shujian Zhang; Wenxuan Zhou; John Lambert; Beck Chen; Bo Li; Andrew Hard; Rajiv Mathews

cs.CLOct 9, 2026

Evo $^2$ : Evolving the Evolver through Weight-Harness Co-Evolution

Authors: Chejian Xu, Shujian Zhang, Wenxuan Zhou, John Lambert, Beck Chen, Bo Li, Andrew Hard, Rajiv Mathews

Organizations: University of Illinois, Urbana-Champaign · Work done as a student researcher at Google DeepMind · Google DeepMind

Abstract

Self-evolving agents are increasingly used to improve language model systems without human intervention through either weight updates or harness improvements. Existing approaches typically pursue these directions separately: some methods update model weights using self-generated experience, while others rewrite the harness that defines how an agent uses prompts, tools, context, and execution feedback. However, these approaches overlook the relationship between model weights and the agent harness. To connect these processes and ensure effective and smooth agent self-evolving, we propose Evo $^2$ , a framework for weight-harness co-evolution. For harness updates, we design a unified dual-role pipeline that a single policy serves as both a Proposer, which writes harness edits, and a Reflector, which diagnoses their effects and guides revision. For weight updates, harness modifications and their execution outcomes guide the training of both the Proposer and the Reflector. We further develop EvoTIDE, a training pipeline that recovers usable trajectories through edit extraction and prefix continuation, then combines SFT with interleaved GRPO across the two roles. Extensive experiments across three coding benchmarks demonstrate that Evo $^2$ significantly outperforms existing self-evolution methods. It achieves 82.1% on Polyglot and 48.3% on SWE-bench Verified, compared with 62.6% and 41.0% for the strongest baseline, respectively. We also show that the evolved harnesses transfer to different tasks without further evolution and rank first in cross-benchmark evaluations. Our dual-role harness improvement framework is also able to extend to self-evolution with proprietary models, improving a GPT-5-mini agent's performance from 24.4% to 48.3%. These results demonstrate the applicability of the framework Evo $^2$ across domains and its broad transferability across models and tasks.

Figures & tables

Figure 1: Weight-harness co-evolution in Evo 2 . A shared Proposer–Reflector policy evolves the agent harness. Harness edits and execution outcomes provide training experience for the policy, whose updated weights guide subsequent search. Crossover searching transfers useful strategies between branches.

Figure 2: Harness evolution and trajectory collection. Parents sampled from the harness pool receive autonomous Type A edits from the Proposer and are evaluated with a fixed task executor. The Reflector uses execution feedback to direct revision or crossover, or to keep the candidate without further editing. Both roles share model parameters, and evaluated candidates extend the pool. The lower panels illustrate prefix continuation, edit extraction, and crossover searching.

Method	Polyglot Pass@1 $\uparrow$	SWE-bench Verified Resolved $\uparrow$	LiveCodeBench Pass@1 $\uparrow$
SICA	60.2	20.5	39.0
DGM	60.2	30.0	54.0
HGM	61.8	41.0	56.0
GEA	62.6	25.0	49.0
Evo 2 (ours)	82.1	48.3	58.0

Table 1: Self-evolution on coding benchmarks. Each column reports the held-out performance of the harness evolved on that benchmark, using Qwen3.8-27B. All scores are percentages; higher is better.

	From Polyglot		From SWE-BenchVerif.		From LCB
Method	SWE	LCB	Polyglot	LCB	Polyglot	SWE
SICA	19.0	33.0	52.0	49.0	43.9	29.0
DGM	20.0	45.0	69.9	45.0	63.4	19.0
HGM	20.0	41.0	53.7	50.0	63.4	16.0
GEA	16.0	45.0	48.8	42.0	61.0	25.0
Evo 2 (ours)	21.0	44.0	72.4	53.0	74.0	43.0

Table 2: Transfer across coding benchmarks, using Qwen3.8-27B. The upper header identifies the benchmark used to evolve and select the best harness; the lower header identifies its evaluation benchmark. Each method contributes one harness per source, evaluated on both other benchmarks without further evolution. Scores use the target benchmark’s metric from Table 1 .

Table 5

Benchmark	Untrained	Trained	$\Delta$ ( $P$ )
Polyglot	61.8	78.0	$+16.3$ ( $.001$ )
SWE-bench Verified	46.0	51.0	$+5.0$ ( $.086$ )
LoCoMo	47.2	56.5	$+9.3$
LongMemEval	57.1	67.7	$+10.6$

Table 5: Effect of weight evolution. Each row compares the untrained evolver (the base Qwen3.8-27B acting as Proposer and Reflector) with the evolver trained by EvoTIDE .

Figure 3: Held-out accuracy versus inference cost. Held-out LiveCodeBench pass@1 of the harnesses in Table 1 , plotted against mean executor completion tokens per held-out task. Evo 2 achieves the highest accuracy, while HGM and DGM use about

1.9\times

as many tokens and

2.6\times

and

3.6\times

as many LLM calls per task, respectively.

Self-Evolving Agents

Self-evolving agents improve their skills, memory, tools, policies, or environments by learning from interaction and evaluation feedback.

See all 427 papers

Sep 29, 2026

Self-Evolving Harness on Multiple Tasks with the Agent as Its Own Optimizer

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 harness obtained at the end of the first stage improves the average score by 4.48 points on the in-distribution benchmarks and by 12.64 points on the out-of-distribution benchmarks, surpassing Codex on the former and matching it on the latter. In the second stage, continued evolution on Claw-Eval, one of the out-of-distribution benchmarks, further raises the score on that benchmark from 66.17 to 68.06, exceeding Codex. We also provide an in-depth analysis of the mechanisms that emerged during evolution, including output truncation, history compaction, and independent review.

Qiankai Xu

Nanjing University

Continual Learning for LLM Agents AI Agent Evaluation Agent Harness Optimization LLM Agent Workflow Optimization LLM Agent Self-Improvement

Ask PDF

Aug 10, 2026

Evo-Bench: Can Language Models Improve Agent Harness?

Large Language Models (LLMs) have driven rapid progress in autonomous agents, yet standard evaluations remain confined to static task solving. An emerging frontier is harness evolution---the agent's capacity to autonomously optimize its own operating harness. However, systematically benchmarking this capability remains challenging, as existing evaluations fail to isolate harness improvements from base model strength, prevent task-specific overfitting, or capture long-horizon iterative research. To address these challenges, we introduce Evo-Bench, the first benchmark designed to evaluate models' intrinsic harness-evolving capabilities across Search, Office, and General agent domains. To rigorously isolate this capability, Evo-Bench employs a novel harness-guided construction framework: it leverages auxiliary-task evolution to identify tasks genuinely sensitive to framework improvements, followed by sensitivity-aware stratified splitting to ensure robust cross-suite generalization. Extensive evaluations across nine frontier and open-weight models reveal that top models achieve massive absolute gains reaching 16.6 points, closely approaching state-of-the-art human-engineered baselines. Crucially, while autonomous evolution outpeforms artificial harness in General tasks and excels in Search tasks, it struggles in Office tasks that demand highly specific processing workflows. Furthermore, our analysis exposes critical temporal anomalies like early saturation, while demonstrating that the synthesized harnesses act as highly transferable reasoning structures, consistently boosting diverse policy models.

Lisheng Huang, Chen Yang, Hao Zhou +6

1Gaoling School of Artificial Intelligence, Renmin University of China · 2BOSS Zhipin, Beijing, China

AI Agent Evaluation Long-Horizon Agent Evaluation Agent Harness Optimization AI Agent Benchmarks

Ask PDF

Aug 5, 2026

EvolveNet: Collaborative Harness Evolution for Agent Self-Improvement

An LLM agent can acquire persistent improvements by evolving its harness---the executable program governing model calls, tool use, verification, and recovery---without updating model weights. When the same harness serves users with different workloads, isolated evolution limits the sharing of useful discoveries, while centralized adaptation can introduce conflicting task-specific changes. We introduce EvolveNet, a framework for collaborative harness evolution that connects local specialization with shared capability accumulation. Clients evolve the same broadcast harness on their local workloads in parallel, and a server composes their adaptations into an updated shared program. Redistribution makes accepted discoveries available to all clients, allowing subsequent local evolution to build on improvements accumulated across the collaboration. EvolveNet supports this process through evidence-guided, scope-typed program aggregation and behavioral validation. Our theoretical analysis provides sufficient conditions for retaining and combining complementary improvements despite program interactions. Experiments across five settings show improved shared-harness performance, with ablation and behavioral analyses supporting the contribution of composition beyond specialist selection alone.

Jun Nie, Yonggang Zhang, Qianshu Cai +3

Hong Kong Baptist University · University of Science and Technology of China · The Hong Kong University of Science and Technology

Agent Harness Optimization Multi-Agent Collaboration LLM Agent Self-Improvement

Ask PDF

Abstract

Self-Evolving Agents

Explore similar work

Self-Evolving Harness on Multiple Tasks with the Agent as Its Own Optimizer

Evo-Bench: Can Language Models Improve Agent Harness?

EvolveNet: Collaborative Harness Evolution for Agent Self-Improvement