Harness-S trains one agent policy on rescues from specialised sibling harnesses
A paper proposes calling a fixed pool of specialised agent harnesses only when the base harness fails, then training the policy on those rescued rollouts so it improves under the original harness.
The problem
In this paper’s framing, an agent policy runs under a harness, the base execution system around it. The authors of Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training (arXiv 2611.04277) start from an observation about harnesses that evolve from a common base. They say this evolution shows Locality, Complementarity, and Nonstationarity: different variants repair different failures, and how useful each variant is changes as the policy learns.
From this the authors conclude that searching for one universally superior harness is the wrong goal. Harness-S instead uses sibling harnesses to selectively rescue failures of a base execution system, then transfers the resulting capabilities into a single policy that stays effective under the original harness.
The paper lists 11 authors. It was published on arXiv on 2026-11-03 and listed on Hugging Face daily papers on 2026-11-02.
How the method works
The abstract describes several parts:
- Directional Harness Evolution develops specialised sibling harnesses along distinct repair directions. Together they form a reusable Sibling Harness Pool, which stays fixed while the policy is trained.
- Complementary Harness Rescue Training calls on this pool only when all initial attempts under the base harness fail. The same policy then generates fresh rollouts under selected auxiliary harnesses, and those trajectories replace the failed group for reinforcement learning.
- An Adaptive Harness Router decides which auxiliary harnesses to use. It selects the Top-k variants using Harness Behavior Profiles and voting across multiple LLMs.
- Policy-aware profile re-anchoring refreshes the profile evidence as the policy learns, since the usefulness of each variant changes during training.
The key design choice is that the extra harnesses are a training aid. At evaluation time the trained policy is returned to the original base harness.
Reported results
The authors evaluate on AppWorld, SWE-Gym, and DBBench. According to the abstract, the trained policy substantially outperforms policy-only training when returned to the original base harness. The authors also say behavioral analyses and held-out cases support the claim that harness-elicited behaviors transfer into the policy itself. Reusing the router for Rank-1 inference gives additional gains at the same executor rollout count. These are the authors’ own claims from the abstract.
The code repository is listed at github.com/DeepExperience/Harness-S.
What the abstract leaves open
The abstract gives no numbers for any of the gains and does not name the base model or the LLMs used for router voting. It does not say how many sibling harnesses the pool contains, what the repair directions are, or what value of k the router uses. It also does not report the training cost of the rescue rollouts and the multi-LLM voting compared with policy-only training, or how often the rescue path is triggered in practice.
Source: Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training (arXiv 2611.04277)