CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?

Jiaqi Su; Guibin Zhang; Jie-Jing Shao; Cong Pang; Tiankuo Yao; Ivor Tsang; Haiyan Yin; Shuicheng Yan

cs.LGOct 18, 2026

CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?

Authors: Jiaqi Su, Guibin Zhang, Jie-Jing Shao, Cong Pang, Tiankuo Yao, Ivor Tsang, Haiyan Yin, Shuicheng Yan

Abstract

Model--harness co-evolution has emerged as a prominent paradigm for recursive self-improving (RSI) agents, jointly evolving their execution scaffolds and intrinsic capabilities. We study the persistent yet often overlooked brain--body mismatch problem in this paradigm, where an evolving harness (body) demands behaviors its model (brain) has not learned or restricts capabilities it already possesses. We introduce CoEvo-RSI, a RSI framework that advances model--harness co-evolution through agent-driven, bidirectional historical inheritance. Drawing on a shared evolutionary archive and execution feedback, a meta-agent coordinates harness development from historical designs with model checkpoint selection or merging and subsequent adaptation under the selected harness, aligning inherited capabilities with evolving execution requirements to reduce brain--body mismatch. Using only supervised fine-tuning (SFT) on its own rejection-sampled trajectories, CoEvo-RSI makes Qwen3.5-9B competitive with larger models across diverse vertical applications, including multi-agent failure attribution, role-aware text-to-SQL, and smart-home instruction following. It surpasses DeepSeek-V4-Pro on TraceElephant (+14.16 agent accuracy) and GLM-5.3 on HomeBench VM (+4.90 success rate), underscoring the need for brain--body alignment and the potential of bidirectional historical inheritance to advance agent co-evolution.

Figures & tables

Figure 2: Brain–body alignment in model–harness co-evolution. Left: Historical harnesses and checkpoints are recombined and adapted to resolve interface mismatch. Right: Step accuracy versus Turn-level Interface Mismatch Rate (TIMR; Section 3.1 ) for ten evolved versions per method, from a three-round run evaluated on the Who&When test benchmark.

Figure 3: Overview of CoEvo-RSI . Left: The Curator draws on archived designs, checkpoints, and outcomes to guide evolution. Center: Harness editors choose historical harness bases; checkpoint selection or merging supplies the paired models. Right: Each round evaluates candidate pairs, tracks TIMR, adapts the selected model to its harness, and promotes improvements by validation performance. Outcomes return to the archive.

Direct LLM attribution
Attribution System	Who&When Hand		Who&When Algo.		Who&When Pro		TraceElephant
Attribution System	Who $\uparrow$	When $\uparrow$	Who $\uparrow$	When $\uparrow$	Who $\uparrow$	When $\uparrow$	Who $\uparrow$	When $\uparrow$
DeepSeek-V4.1-Flash	60.34	12.07	68.25	18.25	47.89	61.77	35.16	23.29
DeepSeek-V4-Pro-0813	60.34	10.34	65.87	22.22	61.77	70.40	56.62	31.96
GLM-5.3	46.55	3.45	45.24	11.11	40.29	52.74	49.32	26.03
MiniMax-M3	53.45	10.34	63.49	17.46	51.26	59.83	52.05	27.85

Table 1: Failure attribution results (%) on the Hand and Algo. subsets of Who&When, Who&When Pro, and TraceElephant. All four benchmarks are held out from training and evolutionary selection.

\uparrow

indicates that higher is better, and “–” marks results not reported. The best/second-best results are highlighted in blue/gray. Results for CoEvo-RSI are averaged over three runs.

Figure 4: Co-evolution on Who&When Algo. Dashed/solid curves show AA/SA, and marker labels give TIMR (%). Panels show harnesses at rounds 6–8. The left panel compares full CoEvo-RSI with Harness-only, Model-only, and Latest-only evolution across ten rounds. The right panels connect selected rounds to actor-summary recovery, budget-aware evidence coverage, and evidence rechecking, including reuse of round-4 model weights. Full CoEvo-RSI improves accuracy while keeping TIMR at or below 0.22%, illustrating the complementary roles of joint adaptation and historical inheritance.

Direct LLM generation
RBAC-Text2SQL System	Spider				BIRD (OOD)
RBAC-Text2SQL System	Safe-EX $\uparrow$	AC-F1 $\uparrow$	Safe-Deny $\uparrow$	Viol. $\downarrow$	Safe-EX $\uparrow$	AC-F1 $\uparrow$	Safe-Deny $\uparrow$	Viol. $\downarrow$
DeepSeek-V4.1-Flash	77.33	94.46	93.80	1.69	62.85	85.97	86.47	8.75
DeepSeek-V4-Pro-0813	78.33	94.90	95.65	2.04	64.16	90.39	90.35	7.00
GLM-5.3	77.41	89.73	79.77	2.89	60.90	82.13	78.29	6.40
MiniMax-M3	76.08	91.01	88.93	5.18	65.29	87.47	86.18	8.93

Table 2: Role-aware text-to-SQL results (%) on the Spider and BIRD subsets of RBAC. OOD marks the subset never used for training.

\uparrow

and

\downarrow

indicate that higher and lower are better, respectively. The best and second-best results are highlighted in blue and gray, respectively.

Figure 5: Harness evolution on role-aware text-to-SQL. Safe-EX and violation rate over ten rounds, with designs from rounds 3, 4, and 9. Blue circles and purple squares track Safe-EX and violation rate on the left and right axes, respectively. Dashed connectors link the selected rounds to evidence-backed refusal, authorization of all SQL references, and verification with bounded repair. Across the run, Safe-EX rises from 77.09% to 80.85%, while violation rate falls from 2.00% to 1.57%, a 21.5% relative reduction.

Direct LLM generation
Home Assistant System	VM		IM
Home Assistant System	Succ. $\uparrow$	F1 $\uparrow$	Succ. $\uparrow$	F1 $\uparrow$
DeepSeek-V4.1-Flash	68.98	86.11	84.54	91.28
DeepSeek-V4-Pro-0813	75.10	89.18	83.51	90.83
GLM-5.3	70.61	86.65	64.95	81.42
MiniMax-M3	54.69	77.79	65.98	84.11

Table 3: HomeBench multi-device instruction following (%, higher is better). VM and IM denote valid and invalid instructions. Blue and gray mark best and second-best results, including ties.

Figure 6: Sensitivity to the number of harness candidates

\mathbf{K}

(three rounds, single run): (a) AA, (b) SA, (c) TIMR, (d) run span, (e) evaluation time, (f) API cost. Each bar corresponds to

\mathbf{K}\in\{4,6,8,10\}

, with the remaining settings fixed. The top row reports held-out task performance and interface violations, while the bottom row summarizes wall-clock run span, evaluation time, and API expenditure. The tested budgets maintain low TIMR with modest variation in accuracy; the default

\mathbf{K}{=}8

gives the shortest run span of 17.48 hours at an API cost of $2.72.

Figure 7: Inference cost–accuracy trade-offs . CoEvo-RSI uses OpenRouter’s Qwen3.5-9B pricing. The left and right panels report Who accuracy on Who&When Hand and Safe-EX on RBAC Spider against the cost of a full evaluation pass, using a logarithmic cost axis. Dashed staircases and shaded regions summarize the baseline Pareto frontiers, while stars mark the co-evolved 9B agent.

Figure 8: Interface compatibility on RBAC. Normalized TIMR ( $\downarrow$ ) across harness–checkpoint pairs; only the lower triangle is available during evolution. Rows index evolved harnesses and columns index model checkpoints, with TIMR globally min–max scaled to $[0,1]$ across the complete matrix for comparison. Stars identify selected pairs and aliased checkpoints, and the side and bottom links record derivation, reuse, and historical design references. Low-TIMR off-diagonal pairings and checkpoint reuse show that interface compatibility depends on the particular harness–model pair, motivating inheritance beyond the latest generation.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

	Attribution	RBAC	HomeBench
Rollout
Max new tokens	8,192	4,096	8,192
Context length	65,536	32,768	65,536
Max agent turns	20	10	20
SFT
Max seq. length	65,536	8,192	65,536

Appendix

Table 4: Per-domain settings. Selection metrics are computed on the validation set following the objectives in the meta-agent prompts ( Appendix E ). AA and SA denote agent and step accuracy.

Split	Composition	#Inst.
Train	TracerTraj 1,618; CORRECT-Error 1,230;	3,400
	AFTraj 430; self-constructed 122
Validation	TracerTraj (held out)	200
Test (OOD)	Who&When Hand	58
Test (OOD)	Who&When Algo.	126
Test (OOD)	Who&When Pro (multi-agent, text-only)	1,750

Appendix

Table 5: Failure attribution data. The 200 validation instances are held out from TracerTraj and excluded from training, and all four test benchmarks are reserved for OOD evaluation.

Split	Source	#Inst.	#DB
Train	Spider (train)	8,000	–
Validation	Spider (train, held out)	2,000	–
Test (ID)	Spider (test)	6,926	20
Test (OOD)	BIRD	10,175	11

Appendix

Table 6: Role-aware text-to-SQL data. Instances are role–query pairs, and BIRD is used only for testing.

Split	#Inst.
Train	10,000
Validation	2,000
Test	342

Appendix

Table 7: HomeBench data. Validation is held out from the official training split.

Method	Sampled	Trained on
GRPO / GSPO (each)	81,600	81,600
CoEvo-RSI (10 rounds)	74,639	39,395

Appendix

Table 8: Training-sample budget on failure attribution. CoEvo-RSI counts are summed over all ten rounds, with replayed trajectories included in the training count.

Shared
	GRPO	GSPO
Learning rate	$1\times 10^{-6}$
Rollouts per prompt ( $n$ )	8
Prompts per step / mini-batch	1 / 1
Epochs (optimization steps)	3 (10,200)
Max prompt / response length	8,192 / 4,096

Appendix

Table 9: Hyperparameters of the RL baselines.

Method	Harness update / inheritance	Model initialization	Model update	Historical weights
				Select	Merge
Model–harness co-evolution
HarnessForge ( Chen et al., 2026b )	Parent harness with archived design exemplars	Parent-branch policy	LoRA SFT	No	No
HarnessX ( Chen et al., 2026c )	Current harness edits; variant branching and routing	Current checkpoint	Cross-harness GRPO	No	No
Co-Harness ( Chen et al., 2026d )	Validated local patches to the current harness	Current checkpoint	SFT	No	No
HASE ( Luo et al., 2026a )	Persistent harness with phase-reviewed edits	Current checkpoint	GRPO	No	No

Appendix

Table 10: Comparison of published evolution procedures. Select denotes choosing model initialization from historical checkpoints independently of the harness’s parent lineage; Merge denotes combining multiple historical model checkpoints. Model updates list the main training instantiations; “–” indicates a fixed model.

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

Recursive Harness Self-Improvement
Jul 17, 2026Hyunin Lee, Jinglue Xu, Jeffrey Seely +3Continual Learning for LLM Agents Agent Harness Optimization
1Sakana AI, 2UC Berkeley
EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution
Oct 7, 2026Python Song, Zhixuan Liang, Kelsey Fu +3Continual Robot Learning Robot Skill Learning
Columbia University · Princeton University
ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement
Sep 14, 2026Siwei Wu, Jincheng Ren, Yizhi Li +12Coding Agents Agent Harness Optimization
University of Manchester · IQuest Research · M-A-P +3

Jul 17, 2026J/K move · Enter open · S save

Recursive Harness Self-Improvement

Hyunin Lee, Jinglue Xu, Jeffrey Seely, Donghyun Lee+2

1Sakana AI, 2UC Berkeley

Continual Learning for LLM Agents Agent Harness Optimization LLM Agent Self-Improvement Prompt Optimization LLM Agent Harnesses

Open paper AskPDF arXiv

Under model--harness co-evolution, harnesses are not merely inference-time scaffolds but data-generating components whose execution traces can shape future foundation models. This motivates harness-in-the-loop learning: optimizing harnesses for both immediate agent performance and the quality of traces used for future model training. However, continually updating provider-built scaffolds is costly and labor-intensive. We therefore investigate whether optimizing user-constructed harnesses in a task-specific manner can improve execution-trace quality while remaining computationally lightweight and requiring only a few update iterations. To this end, we introduce Recursive Harness Self-Improvement (RHI), which represents the harness as a prompt-level specification of the agent loop and iteratively refines it using pairwise feedback over its own revision history. Across 30 synthetic machine-learning research tasks spanning quantitative finance, robotics, and pharmacy, a few RHI iterations suffice to substantially raise the performance ceiling of low-reasoning-effort agents, exceeding the corresponding maximum-reasoning-effort setting while reducing inference cost by up to 60%. We show that these gains arise primarily from improved task-specific context management through more effective inter-agent information flow rather than longer reasoning traces. Finally, we formalize this behavior as an information-theoretic hypothesis for RHI's implicit optimization objective, suggesting RHI as a practical algorithm for continual learning within the paradigm of model--harness co-evolution.

Abstract

Self-Evolving Agents

Explore similar work

Recursive Harness Self-Improvement

EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution

ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement