Harness-S: Generalizable Agent Training by Learning from Sibling Harnesses

Problem

Agent performance depends on the model policy and also on the harness, which bundles prompts, tool interfaces, memory and execution rules. Harness variants derived from the same base fix different failures (locality and complementarity), and their usefulness changes as the policy learns (non-stationarity). This makes it hard to merge them into one universal harness. Evolving the harness and the policy in alternation is costly, and the gains can disappear when the agent goes back to the original harness.

Method

(1) Directional Harness Evolution: the method diagnoses why the base harness H0 fails and creates child harnesses that apply fixes in different directions. All valid variants are kept in a Sibling Harness Pool, which stays fixed during policy training. (2) Complementary Harness Rescue Training: only for tasks where all K rollouts fail under H0, the same policy generates new rollouts with the top-k auxiliary harnesses chosen by a router. These groups replace the failed groups, and training uses GRPO. (3) Adaptive Harness Router: harnesses are chosen using per-harness behavior profiles and votes from several LLMs. The profiles are updated periodically as the policy learns.

Results

• Evaluated under the original H0, the 9B model reached 72.52% AppWorld Pass@3, higher than GRPO (64.89%). It also scored higher on SWE-Gym at 40.00% (GRPO 28.13%) and on DBBench at 76.33% (GRPO 69.67%).
• The 4B model also reached 67.56% AppWorld Pass@3 under H0, ahead of GRPO (60.31%) and H0 rescue GRPO (61.83%).
• It also scored higher than WHALE, a method that co-evolves the policy and harness, which reached 68.70% AppWorld Pass@3 (9B, using its own optimized harness).
• Using the router's rank-1 harness at inference raised the 9B model's AppWorld Pass@3 to 75.19% with the same number of rollouts.
• The authors report that behavior analysis and held-out cases support their claim that behaviors brought out by the harnesses transferred into the policy itself.

Limitations

The experiments cover only two model sizes (4B and 9B) and three benchmarks (AppWorld, SWE-Gym, DBBench). The router needs votes from several LLMs and periodic validation rollouts. Evolving the harness pool in advance also has a cost. The harness pool is fixed during training, so the paper does not cover updating the harnesses themselves as the policy changes.

Why it matters

In practical agent work, tuning a harness often fixes some tasks while making others worse. This approach does not merge several harness improvements into one. It uses them only as a training signal and deploys with the existing harness, which points to a way to reduce the operational burden.

Key terms

Harness

The execution framework around an LLM agent. It includes prompts, tool interfaces, context and memory management, and execution rules.

GRPO

A reinforcement learning method that normalizes the rewards of several rollouts for the same task by the group mean and standard deviation to compute relative advantages. If all rollouts fail, the learning signal disappears.

Rescue Rollout

Retrying tasks that failed every attempt under the base harness with a different harness, to build training groups that contain both successes and failures.

Harness Behavior Profile

A behavior summary, written by an LLM, of the task conditions under which a given harness helps or hinders the current policy.

Original

• arXiv abstract
• PDF
• HTML full text
• Code
• Hugging Face

Authors: Kangning Zhang, Wenxiang Jiao, Bowen Fang, Shuai Shao, Jiachen Zhu, Rong Shan, Shijian Wang, Hao Wang, et al. · Posted 2026-11-03 · License CC BY 4.0 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.