In the improvement of Large Language Model (LLM) agents, prompts, models, or hand-written workflows are typically adjusted, while the execution harness surrounding the model is treated as fixed infrastructure. We argue that this harness itself is a learnable control layer. We formalize harness operation as a finite-horizon Harness MDP, where a lightweight controller selects structural execution actions while the LLM executor remains frozen. The controller is trained from offline rollouts using advantage-weighted regression with only terminal task-rubric rewards.
Additionally, we separate final task quality from a post-hoc Harness Maturity Score, which measures whether the harness follows reliable execution patterns rather than merely whether the final answer is correct. This separation provides a finite-buffer perspective of harness learning: gains in final quality require high-return support in the offline buffer, while process behavior can shift whenever it aligns with advantage-weighted actions.
Across six controlled domains and two public-benchmark adapters, the learned controller consistently enhances verification behavior and selectively improves final task quality, with the largest gains observed on adapted tau-bench retail, adapted AgentBench DB-Bench, and coding with a calibrated structural verifier. Ablations against behavior cloning and Forced CHECK demonstrate that the gains are not merely due to imitation or simply adding checks. These results establish harness control as a learnable layer for frozen LLM agents, while indicating that offline support limits when enhanced process control leads to better final answers.
Blogger's Review: This paper introduces an innovative approach by viewing the execution harness as a learnable control layer, challenging traditional LLM agent architectures. The use of offline reinforcement learning effectively enhances model execution efficiency and task quality, providing new perspectives and possibilities for future research. The practicality and effectiveness of such methods will be tested across various application scenarios.