Meta$^2$: Precomputing Reusable Knowledge for Online Harness Evolution in Long-Horizon Agents

Lingzhi Yuan; Yiguang Cai; Weize Liu; Siqi Yang; Feijiang Han; Yongkyun Lee; Udari Madhushani Sehwag; Furong Huang

cs.AINov 3, 2026

Meta $^2$ : Precomputing Reusable Knowledge for Online Harness Evolution in Long-Horizon Agents

Authors: Lingzhi Yuan, Yiguang Cai, Weize Liu, Siqi Yang, Feijiang Han, Yongkyun Lee, Udari Madhushani Sehwag, Furong Huang

Organizations: University of Maryland · University of Washington · University of Illinois at Urbana-Champaign · All Purpose AI ·

Abstract

Harness optimization improves long-horizon agents by refining the prompts, tools, memory, and hooks around a frozen model. Offline methods can learn from completed trajectories, but deploy one shared harness that may overfit training tasks and cannot react to evidence revealed during a new run. Online evolution can adapt before, during, or after one run through an evolver, but may have to reason about every intervention from scratch or rely on verifier feedback that is usually unavailable at test time. We introduce Meta $^2$ , which moves offline optimization one level up: rather than optimizing a shared harness, it learns reusable knowledge for harness evolution and applies it within each task online. Offline, Meta $^2$ learns a textual Meta-Skill $M$ by contrasting training rollouts of the same task agent under three configurations: without an evolver, with an evolver guided by the current version of $M$ , and with an evolver given privileged verifier information available only in training. Contrasts in trajectories and outcomes are aggregated across tasks to refine $M$ into reusable guidance on when and how to evolve, when to leave the harness unchanged, and which interventions to avoid. Online, an evolver combines the frozen $M$ with emerging trajectory evidence to decide whether to KEEP the harness unchanged or EVOLVE it through a task-local edit. Each edit applies to the rest of the same run without resetting its environment or progress. Thus, evolution guidance transfers across tasks while harness edits remain task-local, and held-out verifier feedback is never available to the evolver. On two long-horizon benchmarks, Terminal-Bench 2.1 and DeepSWE 1.1, Meta $^2$ achieves the highest performance among evaluated methods with GPT-5.6-Terra, reaching 81.00% and 44.44%, respectively. Ablations support the value of contrastive evolution guidance and keeping online harness edits task-local.

Figures & tables

Figure 1: Motivation and overview of Meta 2 . Offline evolution produces a shared harness that may overfit training tasks and cannot adapt to evidence revealed during a new run, while online methods face limitations in intervention timing, knowledge guidance, and verifier availability. By moving offline optimization one level up, Meta 2 learns a reusable Meta-Skill from contrastive training rollouts and freezes it to guide task-local harness evolution at checkpoints online, without access to verifier feedback at test-time. Across both evaluated benchmarks, this design yields the highest performance among the compared methods.

Figure 2: Design of Meta 2 : offline knowledge learning, online harness evolution and format of Meta-Skill. Offline, rollouts are collected under three conditions: the task agent alone (A), evolution guided by the current Meta-Skill (B), and evolution with verifier knowledge (C). Deterministic pairing of terminal outcomes yields act , avoid , and hold contrasts. LLM-based reflection and cross-task aggregation distill these contrasts into reusable lessons, which iteratively update the Meta-Skill. Online, the frozen Meta-Skill guides an evolver LLM to retain or modify the harness at deterministic checkpoints, while the task agent continues along a single trajectory in one environment. Meta-Skill entries specify when, how and why to intervene with reference examples.

Method	Terminal-Bench 2.1				DeepSWE 1.1
	Perf.	Cost ($)			Perf.	Cost ($)
	Perf.	Offline	Online	Full	Perf.	Offline	Online	Full
Human-Designed Harnesses
NexAU Seed	66.33	–	34.09	34.09	34.67	–	62.67	62.67
Codex CLI	68.33	–	56.59	56.59	31.11	–	136.53	136.53
mini-swe-agent	67.33	–	41.08	41.08	29.33	–	68.87	68.87

Table 1: Main performance and cost comparison. Results with GPT-5.6-Terra (medium reasoning) on Terminal-Bench 2.1 (mean@5) and DeepSWE 1.1 (mean@3). Parentheses report the point change from the NexAU seed; costs are aggregate US dollars over each evaluation.

Method	Perf.	Cost ($)
Method	Perf.	Offline	Online	Full
NexAU Seed	70.00	–	16.83	16.83
Self-Evolution Methods
ReasoningBank	69.33 (-0.67)	8.08	21.21	29.29
ACE	65.00 (-5.00)	58.30	21.56	79.86
TF-GRPO	66.33 (-3.67)	37.15	19.86	57.01

Table 2: Robustness across backbones. Terminal-Bench 2.1 test-set results with DeepSeek-V4-Flash-0731 (high reasoning), reported as mean@5. Parentheses report the point change from the NexAU seed; costs are aggregate US dollars over each evaluation.

Figure 3: Checkpoint-interval sensitivity. Performance on Terminal-Bench 2.1 versus log-scaled online cost.

Figure 4: Component ablations. The first four settings evaluate the online evolver and two Meta-Skill construction methods. The fifth retains online edits across tasks instead of resetting to

h^{(0)}

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Meaning
$\xi=(x,\mathcal{E}_{x},V_{x})$	Task instance: public instruction, interactive environment, and verifier.
$s_{t},o_{t},a_{t}$	State, observation, and action at step $t$ .
$\tau_{\leq t}$	Partial trajectory through step $t$ .
$\pi_{\theta}$	Frozen task-agent policy.
$h^{(0)}$	Seed harness.
$h^{(k)}$	Harness after decision $k$ .

Appendix

Table 3: Core notation used in the main text.

Benchmark	Training tasks	Test tasks	Runs per test task
Terminal-Bench 2.1	29	60	5
DeepSWE 1.1	38	75	3

Appendix

Table 4: Task counts and evaluation repetitions for each benchmark.

Figure 5: Early intervention guided by an act entry (Terminal-Bench 2.1). The task requires a Git push to deploy hello.html to a web server on port 8080. At the first checkpoint, the evolver without Meta-Skill specifies an implementation and requests validation only “where feasible.” The task agent tests in a temporary sandbox but does not run the installer in the task environment; the evolver accepts the result despite the missing /git/server directory and inactive service. With Meta-Skill, entry E-02 guides the evolver to require end-to-end evidence: a clone, commit, and push followed by a fresh HTTP request confirming that the deployed page returns “hello world.” The task agent installs the service and verifies this sequence, after which the evolver chooses KEEP at the subsequent and terminal checkpoints.

Figure 7: Avoiding premature redirection with an avoid entry (Terminal-Bench 2.1). By checkpoint

t_{3}

, both task agents have identified the bundled Rave warrior as a promising candidate, but its performance against stone remains below the required threshold. Without Meta-Skill, the evolver uses its final notice to redirect exploration toward other bundled warriors. The task agent constructs an Aeka–Rave hybrid but finishes with 74 wins against stone, one short of the required 75. With Meta-Skill, entry E-06 discourages changing the search strategy while exploration continues to make progress. The evolver chooses KEEP at

t_{3}

and subsequent checkpoints, allowing the task agent to continue tuning Rave and satisfy all five performance thresholds.

	Terra		DeepSeek
Method	Terminal-Bench 2.1	DeepSWE 1.1	Terminal-Bench 2.1
NexAU Seed	$66.33\pm 2.17$	$34.67\pm 1.33$	$70.00\pm 2.64$
Codex CLI	$68.33\pm 5.14$	$31.11\pm 3.08$	–
mini-swe-agent	$67.33\pm 4.18$	$29.33\pm 3.53$	–
Pi agent	$69.00\pm 4.94$	$39.11\pm 6.71$	–
DeepSeek Harness	$68.00\pm 3.42$	$34.22\pm 4.68$	–

Appendix

Table 5: Mean performance

\pm

sample standard deviation across repeated evaluation runs (percentage points). Terminal-Bench 2.1 uses five repetitions over 60 tasks, and DeepSWE 1.1 uses three repetitions over 75 tasks. Terra denotes GPT-5.6-Terra, and DeepSeek denotes DeepSeek-V4-Flash-0731. Dashes indicate settings not evaluated.

Figure 9: Performance versus full cost with GPT-5.6-Terra. Full cost includes one-time offline construction and online evaluation costs, plotted on a logarithmic horizontal scale. Dashed lines connect the non-dominated performance–cost point estimates for Terminal-Bench 2.1 (left) and DeepSWE 1.1 (right).

Method	Offline optimization	Pre-run evolution	Within-run evolution	After-run evolution	Test-time verifier requirement
Meta-Harness [ 20 ]	$\checkmark$	$\times$	$\times$	$\times$	Not required
AHE [ 24 ]	$\checkmark$	$\times$	$\times$	$\times$	Not required
JIT-Agent [ 57 ]	$\checkmark$	$\checkmark$	$\times$	$\triangle$	Mode-dependent
ACE [ 61 ]	$\checkmark$	$\times$	$\times$	$\checkmark$	Mode-dependent
Harness-R1 [ 39 ]	$\checkmark$	$\times$	$\times$	$\checkmark$	Required
Continual Harness [ 18 ]	$\times$	$\times$	$\checkmark$	$\triangle$	Not required

Appendix

Table 6: Evolution regimes and test-time verifier requirements of representative methods.

\checkmark

denotes explicit support,

\triangle

restricted support, and

\times

no support in the described protocol. Verifier requirements refer to deployment-time adaptation rather than offline training or final evaluation.

Panel of Table 7 — Table 7: Evolution regimes by timing, evidence, and update scope. Online edits occur pre-run, within-run, or after-run; after-run updates target either the next attempt or later tasks. Meta 2 learns reusable $\mathcal{M}$ offline and applies task-local edits within-run.

Field	Content
ID / title	A stable identifier for citation and a short title describing the situation.
Type	act supports intervention; hold supports leaving the run unchanged; avoid discourages a particular intervention.
Time point	start , during , or end , indicating the execution stage to which the entry applies.
When to refer	An observable condition that the evolver can assess from the task instruction, trajectory, permitted workspace observations, or channel budget.
Guidance	The harness action to take or avoid, including the channel and message content when applicable.
Why	The pattern observed across training tasks that supports the guidance, including relevant qualifications.

Appendix

Table 8: Fields of a Meta-Skill entry. Each entry specifies an observable situation, evolution guidance, and supporting training examples.

Agent Harness Optimization

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

See all 192 papers

Self-Evolving Harness on Multiple Tasks with the Agent as Its Own Optimizer
Sep 29, 2026Qiankai XuContinual Learning for LLM Agents AI Agent Evaluation
Nanjing University
EvoHarnessBench: Can Your Agents Keep Pace with an Evolving Harness?
Sep 3, 2026Zixuan Ke, Vaidehi Patil, Haizhou Shi +9Continual Learning for LLM Agents LLM Agent Harnesses
Salesforce Research · University of North Carolina at Chapel Hill, conducted this work during an internship at Salesforce Research · University of Wisconsin–Madison +1
LiteEvo: Automated, Cost-Efficient Harness Evolution for Generalization to Unseen Tasks
Sep 27, 2026Euntae Choi, Sumin Song, Sungjoo YooAgent Harness Self-Evolution Agent Harness Optimization
Seoul National University

Sep 29, 2026J/K move · Enter open · S save

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

Qiankai Xu

Nanjing University

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

Open paper AskPDF arXiv

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.

Abstract

Agent Harness Optimization

Explore similar work

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

EvoHarnessBench: Can Your Agents Keep Pace with an Evolving Harness?

LiteEvo: Automated, Cost-Efficient Harness Evolution for Generalization to Unseen Tasks