Harness-S transfers sibling skills into one agent
Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training · arXiv:2611.04277
In plain words
The paper introduces Harness-S, a method for training one software agent to learn from different execution setups. The setups are specialized versions of a shared base: one may help with a task where another fails. The agent tries its usual setup first and uses a sibling setup only when all those attempts fail.
On AppWorld, which tests tasks involving application APIs, the trained agent solved a larger share of test tasks at least once in three tries under its usual setup than a policy-only training comparison. This suggests some rescued behaviors remained after the extra execution support was removed.
The research
For agent training, a failure under one execution setup may reflect how the task is attempted, not only what the shared model can learn.
Harness-S makes it possible to ask whether skills found with specialized setups remain available under the usual one, changing how a training gain should be judged.
Harness-S trains one shared agent from sibling execution setups, using them to rescue failures and then testing whether those behaviors persist under the original setup.
On AppWorld, a test of tasks involving application APIs, evaluators gave the 4B policy three attempts per test task and measured the share solved at least once, averaging three randomized runs. Under the original setup, Harness-S reached 67.56% versus 60.31% for policy-only GRPO.
Because both were tested under that original setup, the gap indicates the gain remained in the trained policy after auxiliary help was removed.
This main comparison covers Qwen3.5-4B on AppWorld; it does not show that every task or randomized run improved. The transfer examples are illustrative, and lasting transfer remains future work.
What the evidence shows
- Why sibling harnesses matter
Harness-S is a method for training one shared agent from specialized sibling harnesses, which are execution variants descended from the same base. The paper identifies three complications: a local edit can fix some failures while regressing on solved tasks; variants can recover one another’s failures; and a variant’s usefulness can change as the policy learns. Together, these observations make a single universally useful setup difficult to rely on.
- Rescue only failed tasks
Harness-S first builds a pool of variants aimed at different failure causes. Validation tasks help select parent variants whose combined coverage reaches different failures, while valid children remain available in the pool.
During training, the current policy uses the original setup first and tries selected siblings only for tasks where all initial attempts fail. This makes harness diversity task-specific and conditional on the policy’s observed failures.
- More mixed training groups
Harness-S produced mixed-outcome groups more often than replacing all-failure groups with fresh attempts under the original harness: 79.38% versus 70.94% of retained groups for Qwen3.5-4B on AppWorld, a test of tasks involving application APIs.
A second measure counted policy tokens in those groups as a share of all retained policy tokens, accounting for trajectory length; it was 80.88% for Harness-S versus 72.68% for GRPO with H₀ rescue. This suggests selective sibling rescue supplied more groups with reward variation for training in this setting; the measures are not final task success.
- Gain remains under the base
AppWorld tests tasks involving application APIs. On its test set, evaluators measured Pass@3, the share of tasks solved at least once in three attempts.
Under the original harness, Harness-S scored 67.56% across three randomized runs, compared with 60.31% for policy-only GRPO. Because the auxiliary harnesses were absent at evaluation, the result supports a retained policy gain on these tasks.
- Transfer has clear limits
Two held-out AppWorld cases show a rescued behavior recurring under the original harness: applying an intersection rule to new song instances and completing account recovery on a different service. In cross-harness tests without further policy updates, Harness-S had the highest Pass@3 among the compared policies across the reported AppWorld and SWE-Gym evaluations, yet remained below DSH H₀ on external harnesses and tests.
The evidence supports transfer in tested cases and settings, with execution support still affecting results.
Deep
- Original paper authors
- Kangning Zhang, Wenxiang Jiao, Bowen Fang, Shuai Shao, Jiachen Zhu
- Submitted to arXiv
- License
- CC BY 4.0