Treat the harness as a trainable control layer with a frozen LLM (Harness MDP)

arXiv 2607.05458·medium signal

arXiv 2607.05458 (submitted July 5, 2026) formalizes agent execution as a finite-horizon 'Harness MDP' where a lightweight controller selects structural execution actions — when to verify, retry, branch — while the LLM executor stays frozen, trained from offline rollouts via advantage-weighted regression using only terminal task-rubric rewards. The paper separates task success from a 'Harness Maturity Score' measuring reliable execution patterns rather than correct answers, and reports gains across six domains including adapted tau-bench retail and AgentBench DB-Bench. Practical route: log rollouts from your existing harness, then train a small controller over structural decisions instead of swapping models or rewriting prompts.