PluginRSI: Recursive Improvement of Agent Harnesses with Reusable Plugins
PluginRSI treats agent harnesses as reusable plugin libraries, not one-off programs to be searched from scratch.
The paper says individual harness mechanisms are optimized separately, stored, and recombined into new harnesses over repeated iterations. It reports gains over existing harness-optimization methods on software engineering, command-line interaction, and question-answering tasks. The resulting harnesses also transfer to other solver models without additional optimization. The authors argue that the accumulated plugin library speeds later optimization and improves convergence on unseen tasks. HF Daily Papers' note
The paper says individual harness mechanisms are optimized separately, stored, and recombined into new harnesses over repeated iterations. It reports gains over existing harness-optimization methods on software engineering, command-line interaction, and question-answering tasks. The resulting harnesses also transfer to other solver models without additional optimization. The authors argue that the accumulated plugin library speeds later optimization and improves convergence on unseen tasks. HF Daily Papers' note
score 5