Harness-S trains AI agents to retain lessons from specialized execution support

An AI agent’s performance depends on more than its model weights. Its harness—the prompts, tool interfaces, memory mechanisms and execution rules surrounding the model—also shapes what it does. A reminder to finish reading every page of results, for example, can change whether an agent gathers all the information a task requires.

In a November 2026 preprint, Kangning Zhang and colleagues introduce Harness-S, a method for turning such execution support into training experience. Rather than searching for one universally superior harness, they preserve specialized variants and use them to revisit tasks that defeat the original system. Their central question is whether the model can retain useful behaviors when those variants are removed.

Keeping different repairs available

The authors describe three patterns in their harness experiments. Edits can repair some failures while breaking previously successful tasks. Different variants can solve different problems. And a variant’s usefulness can change as the underlying model learns.

Harness-S responds by building a pool of related, or sibling, harnesses. An editing process diagnoses failures and develops variants targeting limited groups of related causes. Validation results determine which candidates become parents for further editing, but all executable, protocol-compliant children remain available in the pool.

Concrete edits include reminders to continue pagination until the results are exhausted, instructions to intersect multiple selection conditions, and warnings to reserve an execution step for final submission. These are changes to execution support, not separate specialist models.

The reported routing pool contains the original harness and 19 auxiliaries. Their implementations remain fixed during policy training; the evidence used to select them is refreshed.

Rescue only after every initial attempt fails

For each training task, the current model first makes eight attempts under the original harness. If any succeeds, that group is retained for reinforcement learning.

If all eight fail, a router selects eight distinct auxiliary harnesses. The same model then makes one fresh attempt under each, restarting from the task’s initial state rather than continuing a failed execution. These rescue attempts replace the original failed group for training.

The intended benefit is reward contrast. With the group-relative reinforcement-learning method used here, a group containing both successes and failures provides a signal for reinforcing better actions. Eight uniformly unsuccessful attempts lack that within-task contrast. Rescue creates another opportunity to find successful alternatives without changing the task or its verifier.

The router combines independent rankings from multiple language models. Its evidence includes validation success rates, outcomes on similar tasks, and behavioral profiles describing when each sibling helps or hinders the current policy. Periodic recalibration updates those profiles as the policy changes, without using test outcomes.

Gains after removing auxiliary support

The researchers evaluate Qwen3.5 models with four billion and nine billion parameters on AppWorld, SWE-Gym and DBBench, covering application APIs, software engineering and databases. Main results average three independent randomized runs, each with three execution attempts per test task.

For the four-billion-parameter model on AppWorld, Harness-S achieves 67.56% Pass@3 under the original harness, compared with 60.31% for standard policy-only GRPO training. Pass@3 measures whether at least one of three attempts succeeds. On SWE-Gym, the corresponding scores are 38.75% and 24.38%; on DBBench, 73.33% and 66.67%.

The authors also report improvements over a control that retries failed groups using only the original harness. Selecting one task-specific harness at inference improves the reported means further while keeping the number of executor attempts unchanged. That does not establish equal total computational cost: routing and profile construction require additional work.

Evidence of transfer, with boundaries

Behavioral analysis under the original harness finds more schema consultation and less repeated execution code on AppWorld. Held-out examples show the trained model preserving a set-intersection requirement and completing an account-recovery procedure after related successful rescues entered training updates.

These observations support the authors’ transfer interpretation, but the behavior labels come from an LLM, and the cases are illustrations rather than comprehensive evidence of generalization. The experiments cover two model sizes and three benchmarks; lasting transfer and broader pool reuse remain future research questions.

The method also carries costs. In the reported AppWorld training comparison, Harness-S uses 18.44% more rollouts than plain GRPO, although fewer than the original-harness rescue control. External execution harnesses retain some gains but reduce performance relative to the original setup. The preprint therefore supports a bounded conclusion: specialized execution support can generate useful learning experiences whose benefits persist in these evaluations, without making the surrounding interface irrelevant.