PluginRSI represents an agent's harness as a composition of independent, reusable plugins. It iteratively improves the harness by mutating individual plugins based on execution feedback and then recombining them into new harnesses from a shared library. This approach outperforms existing harness optimization methods on software engineering (SWE-bench), command-line interaction, and question-answering tasks, with a 6–10 percentage point improvement over Meta-Harness on SWE-bench Verified. The resulting harnesses retain their advantage when transferred to other solver models, and the evolved plugin library accelerates subsequent optimization on unseen tasks.
Reusable plugins let agent harness improvements accumulate across tasks
PluginRSI evolves modular agent mechanisms independently, stores successful variants and recombines them into new harnesses.
Top University
Yaorui Shi · Yuchun Miao · Yuxin Chen · Jiayuan Zhang · Yueqing Sun · Xierui Song · +2 more
University of Science and Technology of China · Meituan · Wuhan University · National University of Singapore
Research Digest··2 min read
In one line
PluginRSI improves harness optimization by accumulating and reusing independent plugins, outperforming existing methods across multiple tasks.
Before you read this
Three ideas this paper builds onharness evolutionIteratively adapting an agent's environment, prompts, or evaluation setup to improve task performance and generalization.69 papers →agent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.186 papers →recursive self-improvementAn AI system repeatedly uses its own outputs to enhance its capabilities, creating a feedback loop for autonomous improvement.66 papers →
Read it as
Shi et al. replace whole-program harness rewriting with an optimization loop built around reusable plugins, which encapsulate mechanisms such as roles, skills, memory and tools.
The authors represented an , the software scaffolding that organizes a language model's tools and behavior, as a set of plugins plus a workflow coordinating their execution.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.