HarnessMM: Recursive Self-Improvement for Harness Token Efficiency

Yuhua Jiang; Chang Xu; Feifei Gao; Li Zhao

cs.AIOct 23, 2026

HarnessMM: Recursive Self-Improvement for Harness Token Efficiency

Authors: Yuhua Jiang, Chang Xu, Feifei Gao, Li Zhao

Organizations: Tsinghua University · Microsoft Research Asia (MSRA)

Abstract

Agent harnesses repeatedly invoke reasoning after tool feedback, even when an earlier call has already fixed the next action. We introduce HarnessMM (Harness Minus Minus), a method for improving harness token efficiency through guarded decision reuse and recursive self-improvement (RSI). RSI iteratively refines companion harness configurations from training feedback, with validation selecting the configuration for deployment. A controller at the model-tool boundary reuses resolved decisions to replace later reasoning callbacks while actor weights remain fixed. Executable Decision Graphs store exact actions, literal arguments, causal dependencies, and public guards. Guarded Decision Reuse activates these persistent decisions from observed receipts; Certified Callback Substitution executes them through the native interface and records removal of the registered slow callback. New semantic decisions, unsupported evidence, and selector abstention invoke the original actor again. The representation restricts what feedback may change, while complete-task measurements establish quality and cost. On HumanEval, HarnessMM preserves Codex success at 90.82% while reducing Tokens/task (k) from 66.72 to 47.02 (29.54% saving). For DeepSeek Harness, success rises from 71.43% to 85.71% while Tokens/task (k) fall from 96.28 to 11.62 (87.93% saving). A separate MBPP ablation measures 26.91% lower cost than single-use graphs and 19.52% lower cost than retained callbacks. These results establish guarded decision reuse as a practical way to reduce repeated reasoning in existing harnesses.

Figures & tables

Figure 1: HarnessMM overview. The upper row shows RSI updating a harness from $\mathrm{EDG}_{i}$ to $\mathrm{EDG}_{i+1}$ using task results and token consumption; observations supply execution evidence, including visual signals when supported. Below, GDR introduces guarded reuse of resolved decisions, and CCS replaces a redundant reasoning subgraph with certified native execution. Solid arrows show regular execution, dashed arrows show guarded reuse routes, and dashed boxes mark edited subgraphs. The graphs are schematic.

Panel of Figure 2 — Figure 2: Quality–cost studies of benchmark-specific companion configurations. Quality and actor tokens across six benchmarks for No Skill, SkillOpt, AutoSaddler and HarnessMM. Populations: (a) ALFWorld, 2,133 tasks; (b) SearchQA protected validation set, 3,395 questions (risk-220); (c) LiveMath, 160 questions; (d) AIME, 560 test questions with one execution per method; (e) Terminal-Bench 2 archived partial trials with recorded usage; (f) GAIA2, 480 test scenarios with one execution per method. Qwen3.5-9B serves (a–c), Qwen3.8-27B (d–f). Tokens/task (k) includes actor input and output, including shared authorship; scales differ. Cohort protocols and training budgets appear in Appendix I.1 .

Harness	Pass (%)	Tokens/task (k)	Calls/task	Bypasses
HumanEval ( $n=98$ )
Codex	90.82%	66.72	5.35	0
+HarnessMM	90.82%	47.02	5.09	18
DeepSeek	71.43%	96.28	9.73	0
+HarnessMM	85.71%	11.62	4.23	147
MBPP-sanitized ( $n=257$ )

Table 1: Coding-harness mechanism attribution on HumanEval and MBPP-sanitized with the shared backbone. Calls/task counts model calls; Bypasses counts executed substitutions that skip a model call. Actor costs include input and output tokens.

Panel of Figure 3 — Figure 3: Success and actor token costs for Codex, DeepSeek Harness and their HarnessMM variants. (a,b) use DeepSeek-flash; (e,f) use MiMo-v2.6-pro. Each harness pair shares its backbone, public task prompt, tools and scorer. Costs are reported as Tokens/task (k).

Panel of Figure 4 — Figure 4: Instruction and execution-mechanism ablations. Top: six instruction conditions on the same 2,133 ALFWorld tasks with Qwen3.5-9B; bars average seeds 9401–9403 and whiskers show their range. All conditions use history length two and a 50-step limit; neutral text has the same token count as the task instructions. Bottom: four GDR/CCS execution conditions on the same 257 MBPP-sanitized tasks with the shared backbone and one attempt per task. W/o GDR discards the EDG after one execution; W/o CCS retains the slow callback. W/o EDG execution disables graph use while retaining authorship. Tokens/task (k) include complete actor input and output, graph authorship, control, fallback and unsuccessful executions. Condition definitions and paired uncertainty appear in Appendices I.3 and E .

Figure 5: Eight training-guided harness RSI iterations on MiniWoB++. Each comparison uses the same 32 tasks per arm, Qwen3.5-9B, full screenshots, native rewards, 15 actions and 180 seconds per episode. Calls/task counts physical model requests over all tasks. Tokens/task (k) includes complete episode authorship, control, fallback and unsuccessful calls; optimization and selection costs are separate.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

	Success (%)		Tokens/task (k)		Calls/task
Iteration	Single-use EDG	HarnessMM	Single-use EDG	HarnessMM	Single-use EDG	HarnessMM
1	37.50	34.38	3.91	3.26	3.56	3.00
2	34.38	34.38	3.74	2.95	4.00	3.25
3	90.63	90.63	1.51	1.10	1.47	1.06
4	84.38	84.38	1.88	1.71	2.06	1.91
5	90.63	90.63	1.50	1.10	1.47	1.06

Appendix

Table 2: Complete paired measurements for the eight RSI iterations on all 32 training tasks per arm. Success uses native rewards; tokens include all task executions, including unsuccessful ones.

Arm	Correct (%)	Tokens/task (k)	Bypass/task
Full HarnessMM	93.77	13.17	1.58
W/o GDR	93.00	18.03	0.82
W/o CCS	93.39	16.37	0.00
W/o EDG execution	95.33	16.19	0.00

Appendix

Table 3: Results on all 257 test tasks after the common public-assertion fidelity repair. Every arm includes all 257 tasks and complete actor costs.

Full versus control	$\Delta$ correct (pp)	Cost saving (%)
W/o GDR	+0.78 [-3.50, +5.06]	26.91 [4.01, 45.57]
W/o CCS	+0.39 [-2.72, +3.50]	19.52 [-2.07, 38.68]
W/o EDG execution	-1.56 [-5.45, +2.33]	18.62 [-9.53, 39.78]

Appendix

Table 4: Task-paired differences and full-population cost savings, with Bonferroni-adjusted

98\frac{1}{3}\%

bootstrap intervals over three contrasts. All arms are resampled jointly by independent task for 10,000 draws, seed 20261009.

Arm	Correct (%)	Tokens/task (k)	Bypass/task
Original harness	87.55	67.84	0.00
Full HarnessMM	90.27	16.15	1.79
W/o GDR (the single-use GDR control)	91.44	15.87	0.92
W/o CCS (callbacks retained)	91.83	17.82	0.00

Appendix

Table 5: Results on all 257 test tasks. A bypass is a certified eliminated slow callback. Cost includes all actor input and output over all tasks.

Removal versus full	$\Delta$ correct (pp)	Cost ratio
W/o GDR (the single-use GDR control)	+1.17 [-3.11, +5.45]	0.98 [0.81, 1.22]
W/o CCS (callbacks retained)	+1.56 [-3.11, +6.23]	1.10 [0.90, 1.39]

Appendix

Table 6: Task-paired differences and cost ratios relative to full HarnessMM, with the original five-condition study’s Bonferroni-adjusted

98\frac{1}{3}\%

bootstrap intervals. We resample all 257 tasks jointly across conditions for 10,000 draws, seed 20261007. Cost ratios recompute cohort totals.

Run/arm	SR (%)	Tokens/task (k)	Slow/ep	Fast/ep
Original harness	40.62	891.75	15.69	0.00
CoT-only v20	28.12	1351.62	21.66	0.00
CoT/Jev v20	34.38	1602.29	19.31	8.78
CoT/Jev v22	43.75	753.13	12.62	1.69
CoT/Jev v23	37.50	1071.85	15.72	5.28

Appendix

Table 7: Complete TRAIN costs and scores. Slow and fast calls are counted separately. These are development results.

Condition	Successes /2,133	Success (%)	Tokens/task (k)	Calls/task	Input/task (k)	Output/task (k)
No Skill	902	42.29	34.05	35.61	19.71	14.34
SkillOpt	1,350	63.29	101.33	27.95	94.10	7.23
AutoSaddler	1,317	61.74	40.30	30.02	28.66	11.64
HarnessMM skills	1,412	66.20	31.12	26.43	21.67	9.46

Appendix

Table 8: Fresh, complete ALFWorld instruction-only test with a common actor and environment won-and-done scoring. Each condition is newly executed once per task, with one model call for every action. Tokens include every online input and output, including failed episodes.

Control	Gains	Losses	Success diff. (pp)	95% interval
No Skill	603	93	+23.91	[19.15, 28.34]
SkillOpt	286	224	+2.91	[0.43, 5.61]
AutoSaddler	279	184	+4.45	[2.19, 6.73]

Appendix

Table 9: Secondary paired comparisons of HarnessMM skills against each instruction baseline. Intervals are pointwise, with no multiplicity adjustment. All displayed intervals use 10,000 bootstrap draws over 108 floorplans, retaining all paired tasks within a sampled floorplan; the analysis seed is 20260929.

Condition	Correct/16,980	EM (%)	F1 (%)	Tokens/task (k)
No Skill	12,055	70.9953	79.0006	1.22
SkillOpt	12,986	76.4782	83.2716	3.27
AutoSaddler	12,643	74.4582	81.0417	1.61
Original HarnessMM	13,095	77.1201	83.8780	2.97
Without recovery	13,087	77.0730	83.9784	2.22
With echo rejection	13,108	77.1967	84.0475	2.31

Appendix

Table 10: Complete 9B SearchQA comparison. All online input and output tokens, including drafts, fallbacks and failures, are charged.

SearchQA ( $n=16,980$ )
Condition	Correct	EM (%)	F1 (%)	Tokens/task (k)	Calls
No Skill	12,881	75.86	83.81	1.29	16,980
SkillOpt	13,156	77.48	84.41	3.39	16,980
AutoSaddler	12,639	74.43	81.55	1.76	16,980
Original HarnessMM	13,651	80.39	87.41	2.97	16,980
Without recovery	13,640	80.33	87.30	2.22	16,980

Appendix

Table 11: Complete six-condition QA transfer with Qwen3.8-27B. All responses are fresh, including the archived SkillOpt and AutoSaddler prompt controls. Costs include drafts and fallbacks.

Archived AIME 2025: ledger test set (18, tasks) Qwen3.8-27B
Method	Correct (%)	Tokens/task (k)
No Skill	61.11%	5.60
SkillOpt	72.22%	5.69
AutoSaddler	72.22%	6.81
HarnessMM (ledger)	72.22%	6.79
LiveMath: historical dynamic MCQ adapter, 160 tasks

Appendix

Table 12: Historical benchmark comparisons. Tokens count online input and output; GAIA2 judge usage is separate. Terminal-Bench has method-specific coverage: its token column divides known usage by finished trials, giving lower bounds rather than complete per-task costs. Populations are never pooled.

Condition	Success (%)	Tokens/task (k)
Codex parent	55.56%	2413.56
Codex + built-in compaction	66.67%	2057.72
HarnessMM instruction module	77.78%	1455.83
AutoSOTA-guided 32Ki compaction	61.11%	1068.67
AutoSOTA-guided 96Ki compaction	61.11%	1176.39

Appendix

Table 13: Complete TRAIN development comparison, one execution per task. The instruction is selected from two TRAIN-derived candidates by benchmark performance, then complete-task cost. All selected outputs have executed test execution receipts; failures and compaction calls are included in token totals. This is an instruction/context component study, separate from the CoT/Jev decision mechanism and final held-out evaluation. The last two rows add built-in compaction to the frozen instruction and are rejected because their success counts fall.

Stage	Codex seed	DeepSeek Harness seed
Training: original and two candidates (96 executions)	592.97	862.65
Benchmark judge	0.99	1.05
Two proposals (64 training-scenario exposures)	3.70	2.06
Validation selection (32 executions)	673.91	821.69
Benchmark judge	1.15	1.35
Final original harness (288 executions)	614.78	748.27

Appendix

Table 14: Industrial search and test costs in Tokens/task (k), using all declared executions or training-scenario exposures as the denominator. Cached input is counted once; actor/proposer and benchmark judge costs are separate. DeepSeek actor costs use recorded input plus output; unknown historical request usage is excluded, with all 288 executions retained in each denominator.

Method	Input	Output	Author	Total
No Skill	0.22	4.38	0.00	4.59
SkillOpt	2.27	4.66	0.00	6.93
AutoSaddler	0.29	4.61	0.00	4.90
HarnessMM	0.54	4.28	1.08	5.90

Appendix

Table 15: AIME actor cost components on all 560 test questions, each executed once. Every column uses Tokens/task (k). Input and output are execution components; Author includes the shared selected authorship charge once across the same population. Total sums the unrounded components, including unsuccessful executions.

Method	Input	Output	Total
No Skill	726.45	3.77	730.22
SkillOpt	767.16	3.46	770.63
AutoSaddler	768.38	3.84	772.22
HarnessMM	636.28	3.36	639.64

Appendix

Table 16: Complete GAIA2 actor cost across 480 test scenarios, with one execution per scenario. All columns use Tokens/task (k), including online bank authorship, selection, control, fallback and unsuccessful calls. Native scoring is accounted for separately.

SearchQA: all 3,395 TRAIN questions, Qwen3.5-9B
Policy	Correct	EM/accuracy (%)	F1 (%)	Tokens/task (k)
No Skill	2,404	70.81	78.65	1.22
HarnessMM incumbent	2,616	77.05	83.99	2.29
Target-role guard (640)	2,558	75.35	82.21	0.83
Exact union 640	2,582	76.05	82.77	0.82
Exact union 704	2,588	76.23	82.86	0.87

Appendix

Table 17: Complete training families, with official MRQA EM/F1 for SearchQA and benchmark accuracy for AIME. SearchQA numbers denote source-selection input caps; output remains capped at 512 tokens. All actor input/output, failures and paid recovery are included. SearchQA union variants reuse the recorded controls; AIME freshly executes the official optimized SkillOpt and AutoSaddler policies.

Population	Policy	Success (%)	Tokens/task (k)
Validation: 32 scenarios $\times 3$	No Skill	37.50%	826.77
	Selected HarnessMM incumbent	50.00%	734.32
TRAIN: 32 scenarios $\times 3$	No Skill	30.21%	687.35
	Token-wire v3	9.38%	942.24
TRAIN diagnostic: 8 scenarios $\times 3$	Original harness	50.00%	602.51
	Lazy schemas	25.00%	667.68

Appendix

Table 18: GAIA2 follow-ups with validation, training populations and separately executed gates kept distinct. Actor costs include failures and recovery. The final six rows are complete harness-specific validation diagnostics, not TEST results.

Policy	Success (%)	Tokens/task (k)	Judge/task (k)
No Harness	28.13%	761.25	3.34
Pi (unextended)	20.83%	$[7803.45,7806.14]$	2.50
Official SoL-Pi + original adapter	33.33%	$[2228.57,2230.59]$	2.81
HarnessMM receipt parent (historical)	35.42%	927.03	3.62
HarnessMM exact references (historical)	43.75%	592.65	4.05

Appendix

Table 19: Complete GAIA2 training set (32 tasks) with three benchmark repetitions and common Qwen3.8-27B. Actor input/output includes failures. Bounds reflect five missing shutdown-call usage records, not sampling uncertainty or zero-imputed cost. Official SoL-Pi and Pi share the ARE action bridge; file-edit-specific extensions have no matching tools in this adapter.

SearchQA: training set (3,395, tasks) Qwen3.5-9B
Condition	Correct (%)	Tokens/task (k)
No Skill	70.81%	1.22
Incumbent	77.05%	2.29
risk-220	77.53%	0.97
AIME 2025: training set (6, tasks) Qwen3.8-27B
No Skill	83.33%	6.63

Appendix

Table 20: Complete baseline TRAIN results. Costs include every draft, typed decision, failed call with measured usage and fallback. The starred No Harness row is a historical same-population reference.

Split	Method	EM (%)	F1 (%)	Tokens/task (k)
TRAIN	No Skill	70.81	78.65	1.22
TRAIN	SkillOpt	74.40	80.98	1.61
TRAIN	AutoSaddler	71.13	78.76	1.25
TRAIN	HarnessMM incumbent	77.05	83.99	2.29
TRAIN	HarnessMM risk-220	77.53	84.19	0.97
Validation	No Skill	69.81	77.51	1.23

Appendix

Table 21: Complete SearchQA TRAIN and protected validation, 3,395 questions per split under the frozen 3,395/3,395/10,190 partition and Qwen3.5-9B. Every actor input/output, draft and recovery call is charged. SkillOpt and AutoSaddler use their pinned official optimization engines with the same full-context MRQA adapter. The final TEST split is sealed.

Seed	Condition	Success (%)	Tokens/task (k)	Calls/task
9401	HarnessMM	66.20	31.12	26.43
9401	w/o general guidance	61.84	28.75	28.70
9401	w/o task-family guidance	49.27	39.04	33.54
9401	w/o skills	42.29	34.05	35.61
9401	w/ mismatched task family	48.48	43.03	34.11
9401	w/ neutral text	41.07	43.34	36.02

Appendix

Table 22: Complete benchmark-audited instruction ablations. Each row contains all 2,133 tasks. Replication seeds reuse task identities and are not independent task populations. Only complete six-condition seeds are shown.

Seed	Control	Full minus control (pp)	95% interval	Gains	Losses
9401	w/o general guidance	+4.36	[2.61, 6.21]	232	139
9401	w/o task-family guidance	+16.92	[13.16, 20.28]	470	109
9401	w/o skills	+23.91	[19.15, 28.30]	603	93
9401	w/ mismatched task family	+17.72	[14.15, 21.02]	491	113
9401	w/ neutral text	+25.13	[20.32, 29.44]	621	85
9402	w/o general guidance	+2.11	[0.48, 3.76]	215	170

Appendix

Table 23: Paired benchmark-success differences against the full skill. Pointwise 95% intervals use 10,000 bootstrap draws over 108 floorplans, with analysis seed 20260930. All paired tasks in a sampled floorplan stay together. The exploratory intervals have no multiplicity adjustment.

Configuration	Success (%)	Tokens/task (k)	Calls/task
Rules + actor	92.82	1.08	1.04
Rules only	92.39	0.00	0.00

Appendix

Table 24: All 710 validation tasks, seed 9401, common 50-step limit, public observations and tool actions. Failures remain in both denominators. Zero model tokens do not imply zero CPU computation.

Agent Harness Optimization

Agent harness optimization improves the prompts, tools, memory, workflows, and control code that shape how AI agents act.

See all 183 papers

Oct 19, 2026

HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs

Harness development plays a crucial role in building capable AI agents. Recent approaches improve harnesses through recursive self-improvement (RSI), but their performance gains often come with higher execution costs. We introduce HarnessPro, a label-free method for evolving harnesses to improve task performance while reducing execution cost. In this setting, candidate harnesses are evaluated using LLM-generated tests without access to ground-truth solutions or official tests. Specialized agents diagnose execution trajectories, plan updates with different performance-cost priorities, and implement them as candidate harnesses. Rather than returning a single harness, HarnessPro uses normalized weighted squared-distance selection to provide cost-focused (lite), balanced (medium), and performance-focused (heavy) harnesses for different deployment budgets. Selected harnesses enter the next iteration, where execution feedback guides further updates. Our experiments cover Terminal-Bench 2, SWE-Bench Pro, and ProgramBench, using GPT-5.5 with Codex and GLM-5.2 with Claude Code. Across these benchmarks using GPT-5.5 with Codex, the heavy mode of HarnessPro achieves an average pass-rate gain of 2.0 percentage points and an average per-task cost reduction of 26.1% relative to RHO. This advantage also holds when using GLM-5.2 with Claude Code. Code is available at https://github.com/pzs19/HarnessPro

Zhuoshi Pan, ZhiYuan Feng, Yukuo Cen +6

Tsinghua University · Z.AI Co., Ltd. · Shanghai Jiao Tong University +2

Agent Harness Optimization LLM Agent Harnesses LLM Agent Self-Improvement

Ask PDF

Oct 4, 2026

MESH-Harness: Self-Improving Agent Harnesses via Bandit-Guided Compositional Evolution

An agent harness is the code that organizes context, maintains state, and coordinates tool calls for a language model. We study how to improve the harness under a limited evaluation budget while keeping model weights fixed. Our method, MESH-Harness, organizes each harness into functional modules with explicit role-specific interfaces, allowing alternative implementations of each module to be substituted and recombined. It uses shared module representations and full-covariance LinUCB to score candidate combinations based on predicted performance and exploration value. Mixed-start coordinate ascent selects complete configurations for evaluation without enumerating the combinatorial space. Validation traces then guide local code edits, and the resulting candidates are incorporated into fixed-capacity role-specific pools for subsequent recombination. On text tasks, retrieval-augmented mathematical reasoning, code generation, and interactive scientific tasks, MESH-Harness outperforms Meta-Harness by 5.70, 7.01, 2.00, and 5.00 points, respectively, under matched candidate-evaluation budgets. Iterative harness optimization improves MESH-Harness by 5.63-7.79 points over its first-round configurations. For the reported configurations, aggregate test-time cost is 44.2% lower than that of Meta-Harness, while total cost including search is 14.6% lower. These results show that combining module-level design reuse with feedback-driven compositional search can systematically improve agent harnesses while keeping overall optimization cost under control.

Zhiwei Shang, Yu Huo, Mingrong Gong +6

The Chinese University of Hong Kong, Shenzhen · DeepWisdom · The Chinese University of Hong Kong +3

Agent Harness Optimization LLM Agent Workflow Optimization LLM Agent Self-Improvement LLM Agent Harnesses

Ask PDF

Jul 17, 2026

Recursive Harness Self-Improvement

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.

Hyunin Lee, Jinglue Xu, Jeffrey Seely +3

1Sakana AI, 2UC Berkeley

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

Ask PDF

Abstract

Agent Harness Optimization

Explore similar work

HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs

MESH-Harness: Self-Improving Agent Harnesses via Bandit-Guided Compositional Evolution

Recursive Harness Self-Improvement