PluginRSI: recursive harness improvement via reusable plugins beats whole-program search

Yaorui Shi · hf · 2026-10-06

Core idea: the harness around an LLM is a central determinant of agent performance, but existing methods search over complete programs where individual mechanisms are hard to isolate and reuse.
PluginRSI represents a harness as a composition of atomized plugins: plugins are improved independently, accumulated in a shared library, and recombined into new harnesses each iteration.
Results: outperforms existing harness optimization methods on software engineering, command-line interaction, and QA tasks; evolved harnesses transfer to other solver models without further optimization; the plugin library accelerates subsequent optimization and speeds convergence on unseen tasks.
Conclusion: accumulating reusable mechanisms provides an effective basis for continued harness improvement.

More from coding & agent