RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

A method called RRSI (Regularized Recursive Self-Improvement) has been proposed to automate the improvement of LLM agent harnesses — the prompts, control flow, tooling, memory, and context management surrounding a frozen backbone model — by iteratively proposing and selecting component-wise edits. The approach targets the harness layer rather than the underlying model, which the work states largely magnifies an agent's capability.

An LLM agent's capability is largely magnified by its harness, namely the prompts, control flow, tooling, memory, and context management surrounding the frozen backbone model. Recent methods increasingly automate this process by iteratively proposing and selecting component-wise edits of an agent ha