ModularRSI splits agent harness into five modules, claims transfer

Recent recursive self-improvement work on agent harnesses tends to boost benchmark scores without producing scaffolding that survives a change of task or model. In a paper posted to arXiv, a group led by Siwei Wu proposes fixing that by taking the harness apart.

The framework, ModularRSI, "decomposes the evolvable harness into five functional modules: Agent Loop, Tool Use, Observation Management, Context Management, and Task Completion Detection." Each module evolves independently within a restricted modification scope, and an integration stage then combines the evolved modules and resolves conflicts. To keep the process from silently overfitting to the evaluation set, the authors curate "2,000 executable evolution tasks from external sources that are disjoint from downstream evaluation benchmarks."

The headline result is qualitative rather than tabular. "Experiments on TB2.0 and SWE-Bench Verified show consistent improvements on unseen in-domain and cross-domain tasks, with the evolved harness also transferring across different foundation models," the paper says.

The abstract publishes no per-benchmark scores, no baseline comparisons, and does not name the foundation models on which cross-model transfer was tested.