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

Harness-S. Learning from Sibling Harnesses for Generalizable Agent Training. By Kangning Zhang and ten others, from Shanghai Jiao Tong University, Xiaohongshu, and three other institutions. Published on the third of November.

An agent's harness is the execution system wrapped around its policy, and harnesses get evolved: variants branch off a common base. The authors observe three properties of that evolution. Locality, because different variants repair different failures. Complementarity, because those repairs do not overlap. And Nonstationarity, because a variant's usefulness shifts as the policy itself learns. Taken together, these suggest that hunting for one universally superior harness is the wrong goal.

Harness-S takes a different route. It keeps the original base harness as the place the policy must ultimately perform, and uses sibling harnesses only to selectively rescue that base system's failures. Whatever capability the siblings elicit is then transferred into a single policy, one that stays effective once it is returned to the original harness alone.

Building the siblings is the job of Directional Harness Evolution. It develops specialised sibling harnesses along distinct repair directions, and collects them into a reusable Sibling Harness Pool that stays fixed while the policy trains. Training itself is Complementary Harness Rescue Training. The pool is called on only when every initial attempt under the base harness fails. In that case the same policy produces fresh rollouts under selected auxiliary harnesses, and those trajectories replace the failed group for reinforcement learning.

Choosing which siblings to use falls to an Adaptive Harness Router. It picks the Top-k variants using Harness Behavior Profiles together with voting across multiple LLMs. Because a sibling's value drifts as the policy improves, policy-aware profile re-anchoring refreshes that evidence during learning.

The method is tested on AppWorld, SWE-Gym, and DBBench. On all three, the trained policy, once returned to the original base harness, substantially outperforms policy-only training. Behavioural analyses and held-out cases are offered as support that the behaviours the harnesses elicited have moved into the policy itself, rather than staying in the scaffolding. There is a further use for the router at inference time: reusing it for Rank-1 inference yields additional gains at the same executor rollout count.

Siblings serve here as tutors, not replacements. They step in only when the base harness has failed every attempt, their successful trajectories become training signal, and then they step away. What remains afterwards is one policy, run under the harness it started with, carrying behaviours it first learned under someone else's scaffolding.