Harness-S: Leveraging Sibling Harnesses for Generalizable Agent Training

In agent training, the design of the execution harness often determines the final performance of the policy. However, a common dilemma is that a harness optimized for a specific failure mode may gradually become ineffective as the policy learns. The recently proposed Harness-S method offers a new approach to this challenge through the collaborative training of "sibling harnesses."

Three Key Properties of Harness Evolution

The research points out that variants evolved from the same base harness exhibit three key properties: Locality, Complementarity, and Nonstationarity. Locality means that different variants fix different failure modes; complementarity indicates that these fix directions can complement each other; and nonstationarity points out that the effectiveness of a harness changes dynamically as the policy learns.

Together, these three properties reveal a fact: there is no universally optimal harness. Instead, the value of a harness is context-dependent and evolves over time. Therefore, rather than pursuing a "one-size-fits-all harness," it is better to build a training mechanism that can dynamically leverage the strengths of different harnesses.

Core Mechanisms of Harness-S

Harness-S aims to selectively leverage sibling harnesses to rescue failures of the base execution system and transfer the resulting capabilities into a single policy, so that it remains effective under the original harness. The method consists of three key components:

Directional Harness Evolution

This component develops specialized sibling harnesses along different fix directions, forming a reusable Sibling Harness Pool. This pool remains fixed during policy training, ensuring stability of the training process.

Complementary Harness Rescue Training

This harness pool is invoked only when all initial attempts under the base harness fail. The same policy generates new trajectories under the selected auxiliary harness, and these trajectories replace the failed groups for reinforcement learning. This "on-demand rescue" mechanism avoids unnecessary computational overhead while ensuring the quality of the training signal.

Top-k selection means that the router does not pick just one sibling harness, but ranks candidates by confidence or vote score and takes the top k. k is typically small (e.g., 2–3) to control the additional rollout overhead during the rescue phase. Multiple candidates can cover different failure modes, increasing the probability of successful rescue.

Multi-LLM voting is an ensemble-style routing approach: several large language models each judge, based on harness behavior profiles, which sibling harness is most suitable for the current failure sample, and then aggregate via majority vote or weighted scores. Compared with a single model's judgment, voting reduces routing noise but introduces inference cost, so it is typically used only during training or in latency-insensitive inference scenarios.

Adaptive Harness Router

The router uses Harness Behavior Profiles and votes from multiple large language models to select the Top-k variants. As the policy learns, policy-aware profile re-anchoring refreshes this evidence, ensuring that routing decisions remain consistent with the current state of the policy.

Harness in the agent domain refers to a runtime shell built around a base model, responsible for tool calls, context management, error recovery, retry strategies, and other interaction logic. It does not change model parameters, yet it directly determines what the agent can observe in the environment and which actions it can take. The same base model paired with different harnesses can differ in task success rate by tens of percentage points, so harness design itself is an optimizable "system-level policy."

Policy in this context is the agent model being trained, which outputs actions based on observations. The goal of reinforcement learning training is to make the policy maximize cumulative reward within the interaction loop of a given harness.

"Groups" and trajectories in reinforcement learning (RL): methods like GRPO sample multiple trajectories for the same prompt to form a group, and estimate gradients using relative advantages within the group. Harness-S's "rescue" precisely replaces trajectories in failed groups, thereby improving the training signal without increasing the additional rollout budget.

Experimental Validation and Transfer Effects

On three benchmarks—AppWorld, SWE-Gym, and DBBench—the trained policy, when returned to the original base harness, significantly outperforms the baseline of policy-only training. Behavior analysis and held-out cases further support that the behaviors induced by the harness indeed transfer to the policy itself.

Furthermore, reusing the router for Rank-1 inference provides additional performance gains under the same number of executor rollouts. This shows that Harness-S's mechanisms are not only applicable during training but can also bring gains at inference time.

Rank-1 inference means that at inference time, the router selects only the single highest-scoring sibling harness to execute the task, rather than rolling out directly on the original base harness. Since the router reuses the harness behavior profiles learned during training, this switch adds almost no extra overhead; under the same number of executor rollouts, Rank-1 inference effectively treats "choosing the right harness" itself as a free performance gain.

Held-out cases refer to task instances not seen during training. Evaluating on them can test whether the policy has truly acquired transferable capabilities rather than memorized trajectory patterns under a specific harness. Behavior analysis typically involves comparing the policy's action distributions and error types under different harnesses to determine whether the behavior changes induced by the harness have been internalized into the policy.

Implications and Outlook

Harness-S offers a pragmatic approach: rather than pursuing a single optimal harness, build a harness ecosystem and dynamically exploit its diversity during training through intelligent routing. This approach avoids the "one-size-fits-all" problem in harness design while maintaining the generality of the policy.

For the field of agent training, this work highlights the coupling between harness evolution and policy learning. Future research can further explore how to automatically discover new fix directions and how to extend this mechanism to more diverse task environments.

---

References

Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training. (Compiled from the provided source material)