The paper proposes ModularRSI, a framework for evolving AI agent execution harnesses (coding/terminal agents) that decomposes the harness into five independently evolvable modules (Agent Loop, Tool Use, Observation Management, Context Management, Task Completion Detection) to avoid the entanglement problems of whole-harness optimization. It uses contrastive analysis of successful versus failed trajectories across 2,000 benchmark-disjoint evolution tasks to identify recurring behavioral deficiencies rather than benchmark-specific fixes. Experiments on TB2.0 and SWE-Bench Verified show consistent gains on unseen in-domain and cross-domain tasks, with the evolved harness transferring across different foundation models.
