Friday, 6 November 2026 · 8 min read
Harness-S improves agent policies by learning from complementary execution variants
Harness-S selectively uses sibling harnesses to rescue failures of a base execution system, transferring their complementary behaviors into a single policy that performs better under the original harness. It outperforms policy-only training on multiple benchmarks and model sizes.
Based on the paper Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training by Kangning Zhang, Wenxiang Jiao, Bowen Fang, Shuai Shao and 7 others
Challenges in evolving execution harnesses for agent policies
An agent's behavior depends not only on its policy model but also on its execution harness, which includes prompts, tool interfaces, context, and execution rules. Different harness designs can significantly affect task success even with the same policy.
Harness evolution creates sibling harnesses by modifying the base harness to repair failures. However, these variants exhibit locality (each fixes some failures but may regress others), complementarity (different variants fix different failures), and nonstationarity (their utility changes as the policy learns).
These properties make it difficult to combine all repairs into a single universal harness, and joint harness-policy evolution is costly and may not transfer improvements back to the base harness. This motivates a method to selectively use sibling harnesses to rescue failures and transfer their benefits into a single policy effective under the original harness.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Two figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short