CHART: A Harness-Rotation Curriculum for Harness-Robust Search Agents

Xinlu Zhang; Ying-Chun Lin; Zhihan Zhang; Besnik Fetahu; Xi Chen

cs.LGSep 5, 2026

CHART: A Harness-Rotation Curriculum for Harness-Robust Search Agents

Authors: Xinlu Zhang, Ying-Chun Lin, Zhihan Zhang, Besnik Fetahu, Xi Chen

Organizations: Amazon

Abstract

Search agents are usually trained under a single harness. But once an agent is deployed in a real application, its harness is frequently updated (e.g., a rewritten system prompt) to fit production needs. This exposes a fragility of post-trained agents: because a learned behavior is entangled with its training harness, even a harness update that leaves the task unchanged can fail to elicit the behavior. We train a search agent to perform parallel search, a popular strategy for improving both search efficiency and performance. We find that training under a fixed harness makes the behavior harness-local, overfit to that harness's surface form: when the harness changes, the model falls back to serial search. An intuitive fix is harness augmentation, but simply training on more harnesses does not resolve the problem. GRPO learns from the reward gap between parallel and serial rollouts of the same question: a small harness pool saturates that gap early, while a large pool dilutes the per-harness signal too thinly for any harness to consolidate. We therefore propose Curriculum HArness Rotation Training (CHART), a rotating curriculum that lets a search agent gradually consolidate parallel search across harnesses. At each periodic evaluation, CHART "graduates" the harnesses whose expected behavior is learned and replaces them with still-learnable ones, keeping the reward gap alive throughout training. Starting from the same harness pool, CHART makes the model learn parallel search on all harnesses, whereas static augmentation succeeds on at most half of them. The behavior also carries to held-out harnesses: CHART parallelizes on 89% of held-out turns, against at most 5% for the static pools. It further transfers to a new QA task and search environment, improving pass@1 by 5.6pp over the best static pool. Finally, CHART-trained agents benefit more from meta-harness search than baselines.

Figures & tables

task	harn. $\times$ data	role
in-pool	$33{\times}150$	all 33 harnesses, incl. untrained
held-out	$12{\times}150$	12 unseen harnesses
shopping	$19{\times}200$	new task $+$ search env

Table 1: Eval tasks; each reports pass@1 and

\mathrm{PR}\%

	in-pool (33 harnesses)					held-out harnesses		shopping (OOD)
		$\mathrm{PR}\%$
Model	pass@1	cue	cue-free	mean	cov.	pass@1	$\mathrm{PR}\%$	pass@1	$\mathrm{PR}\%$
base (no train)	0.364	37%	5%	24%	3/33	0.358	2%	0.375	5%
fixed pool of $N$ harnesses
fixed-1	0.456	45%	2%	27%	6/33	0.406	0%	0.381	11%
fixed-3	0.507	67%	8%	42%	9/33	0.496	2%	0.381	17%

Table 2: All models at 200 training steps on the three validation tasks of § 4 ; no model trains on the held-out or shopping harnesses.

\mathrm{PR}\%

is a per-harness mean, split in-pool into cue and cue-free (§ 4 ), with mean over all 33. cov. is coverage (

\mathrm{PR}\%\geq 0.80

, § 3 ). CHART stops at

N{=}9

since fixed-33 is also CHART-33 . Bold and underline mark the best and second-best pass@1; the

\mathrm{PR}\%

columns are read as the cue-vs-cue-free gap, not as a maximum.

task	$\mathrm{PR}\%$ 200 $\to$ 260	pass@1 200 $\to$ 260
shopping (OOD)	59 $\to$ 100 ( $+41$ )	.437 $\to$ .491 ( $+5.4$ )
in-pool	85 $\to$ 100 ( $+15$ )	.512 $\to$ .534 ( $+2.2$ )
held-out	89 $\to$ 100 ( $+11$ )	.513 $\to$ .527 ( $+1.4$ )

Table 3: CHART-3 trained 60 steps past the shared budget (200

\to

260), shown as start

\to

end with the gain in parentheses (pp); pass@1 omits the leading zero (.437

=

0.437). pass@1 rises most where

\mathrm{PR}\%

had the most room to grow.

		pass@1		$\mathrm{PR}\%$
Task	Model	best-of-search	top-5 mean	best-of-search	top-5 mean
shopping	base	0.415	0.451	100%	99%
	CHART-1	0.445	0.464	100%	100%
	CHART-3	0.515	0.485	100%	100%
deep-research	base	0.409	0.387	6%	19%
	CHART-1	0.473	0.477	90%	91%

Table 4: Meta-harness search over three training levels: a more-trained substrate searches to a better harness on both tasks and metrics. deep-research pass@1 and

\mathrm{PR}\%

are computed over the pooled FRAMES

+

seal_hard instances (

200{+}254

); per-set breakout in Appendix D .

Model	base champ	CHART-1 champ	CHART-3 champ	row mean	row spread
	harness (each model’s val champion)
base	0.418	0.316	0.347	0.360	0.102
CHART-1	0.449	0.429	0.431	0.436	0.020
CHART-3	0.490	0.460	0.450	0.467	0.040

Table 5:

3\times 3

training

\times

harness cross on the shopping test set,

n{=}8

samples per cell: more training both helps—row mean rises, every column monotone base

<

CHART-1

<

CHART-3—and flattens harness-sensitivity, as row spread shrinks. Best per column in bold.

Table 6

Table 7

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Strategy	pass@1	turns	LLM calls	searches	input tok	output tok
Sequential ( $P{=}1$ , $T{=}25$ )	0.473	10.9	11.9	10.9	67.0k	2.1k
Parallel (fixed-1, $P{=}5$ , $T{=}5$ )	0.513	2.5	3.4	10.5	26.5k	2.2k

Appendix

Table 10: Budget-matched single-harness comparison (deep-research val, 150 prompts,

N{=}1

). Same base model and

\sim

25-search budget; parallel is more accurate and far more efficient. Means are per prompt.

Test set	base	CHART-1	CHART-3
FRAMES	0.550	0.645	0.705
seal_hard	0.268	0.350	0.378
overall (pooled)	0.409	0.473	0.535

Appendix

Table 11: Deep-research best-of-search pass@1, per test set. The two sets are pooled at the instance level (

200{+}254

questions) into the deep-research pass@1 row of Table 4 ; because pooling weights by set size, the overall row is not the simple average of the two. Both rise monotonically base

\to

CHART-1

\to

CHART-3.

LLM Agent Training

Large language model agent training teaches models to plan, use tools, and learn from multi-step interaction with environments.

See all 590 papers

Oct 4, 2026

Harness-Search: Guiding Long-Horizon Search through Multi-Agent Coordination

Long-horizon search requires agents to gather evidence across multiple steps and synthesize it into well-supported answers. The recent agent harnesses provide a natural and promising framework to support such long-running search processes. As interaction histories grow, one single agent in harnesses might get stuck and cause the policy to lose track of unresolved questions, overlook useful evidence, or terminate before sufficient support has been collected. One of promising way is to decouple three distinct responsibilities of proposing retrieval actions, updating persistent state, and deciding when to stop rather than concentrating them within a single policy. Targeted at it, we introduce Harness-Search, a multi-agent search harness to reduce the local errors propagating across subsequent exploration, evidence curation, and termination decisions. In particular, Harness-Search assigns these responsibilities to three permission-bounded authorities: a Retrieval Policy that proposes search operations, a Memory Operator that validates and commits persistent-state updates, and a Summary Auditor that accepts or rejects termination based on the sufficiency of the curated evidence. Together, these roles form a Propose-Commit-Audit loop in which actions are proposed, persistent evidence is selectively committed, and stopping decisions are subjected to an explicit sufficiency check. Across seven long-horizon search benchmarks, Harness-Search improves both retrieval and answer generation under the same policy backbone, increasing Recall by 4.60-27.92 points and Final-Answer Recall by 12.34-30.13 points over the strongest harness-based baseline on each evidence-retrieval benchmark. Moreover, trajectory-level analyses show that Harness-Search continues to accumulate useful evidence and expand evidence coverage with less redundant retrieval as the search history grows.

Shanyong Wang, Zhenwen Ji, Lei Jin +7

Xiaohongshu Inc. · the Joint SDU-NTU Centre for Artificial Intelligence Research (C-FAIR), Shandong University · University of Illinois at Urbana-Champaign

Multi-Agent Coordination Multi-Agent Systems Memory-Augmented Agents LLM Agent Harnesses Long-Horizon Reasoning

Ask PDF

Jun 1, 2026

Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses

Search agents are often trained as policies over growing transcripts: the model must decide how to search while also remembering what it has seen, which evidence is useful, which constraints remain open, and which claims have actually been checked. We argue that this formulation puts too much routine state management inside the policy: reinforcement learning is forced to optimize both semantic search decisions and recoverable bookkeeping that the environment can maintain more reliably. We introduce Harness-1, a 20B search agent (retrieval subagent) trained with reinforcement learning inside a stateful search harness. The harness maintains environment-side working memory, including a candidate pool, an importance-tagged curated set, compact evidence links, verification records, compressed and deduplicated observations, and budget-aware context rendering. The policy retains the semantic decisions: what to search, which documents to keep or discard, what to verify, and when to stop. Across eight retrieval benchmarks spanning web, finance, patents, and multi-hop QA, Harness-1 achieves 0.730 average curated recall, outperforming the next strongest open search subagent by +11.4 points and remaining competitive with much larger frontier-model searchers. Its gains are especially strong on held-out transfer benchmarks, suggesting that reinforcement learning over explicit search state can produce retrieval behaviors that generalize beyond the training domains. Our code is available at https://github.com/pat-jj/harness-1.

Pengcheng Jiang, Zhiyi Shi, Kelly Hong +5

University of Illinois at Urbana-Champaign · †UC Berkeley · ✧Chroma

Reinforcement Learning Agentic Retrieval RL for IR RL for LLM Agents Memory-Augmented Language Model Agents

Ask PDF

Oct 15, 2026

Harness-Delta Attribution: Dissecting Gains From Harness Evolution

Harness evolution automatically optimizes the prompts, skills, tools, and orchestration code around a fixed-weight model, offering a way to improve agentic systems without updating model weights. The agent harness is iteratively optimized to achieve higher performance on the search set. But how much of this gain reflects benchmark-specific overfitting or additional inference-time compute, and how much may generalize beyond the search set? In this work, we introduce Harness-Delta Attribution (HDA), an evaluation protocol that dissects an evolution gain into artifact-driven overfitting, gains explained by test-time scaling, and the remainder, which gives an upper bound on residual generalizable improvement. We apply HDA to 11 executor models and three harness-evolution frameworks across math, coding, creativity, and agentic tasks. In 21 of these settings that show statistically significant search-set gains, we find that a large share of the improvement is explained by overfitting or test-time scaling. Specifically, among these 21 settings, overfitting and test-time scaling account for 54% and 14% of the gain on average; their combined share exceeds 50% in 14 settings. The residual and held-out transfer vary substantially across tasks, models, and frameworks. Importantly, search-set gains often fail to transfer to held-out tasks, and provide limited evidence that they will persist under practical distribution shifts. Moreover, validation-based signals are insufficient to prevent overfitting. These results motivate attribution-aware evaluation in future harness evolution research.

Wenxuan Ding, Qin Lu, Changlong Yu +5

Amazon · New York University

Agent Harness Optimization Agent Evaluation Benchmark Contamination Test-Time Scaling

Ask PDF

Abstract

LLM Agent Training

Explore similar work

Harness-Search: Guiding Long-Horizon Search through Multi-Agent Coordination

Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses

Harness-Delta Attribution: Dissecting Gains From Harness Evolution