Regularized Recursive Self-Improvement of Agent Harnesses (RRSI)

RRSI is a framework for the recursive self-improvement of LLM agent harnesses, designed to optimize prompts, control flow, tools, and memory. It addresses overfitting in agent evolution by regularizing both the proposal and selection processes, ensuring robust performance across diverse task domains.

RRSI is a framework for the recursive self-improvement of LLM agent harnesses, designed to optimize prompts, control flow, tools, and memory. It addresses overfitting in agent evolution by regularizing both the proposal and selection processes, ensuring robust performance across diverse task domains.

• 01Regularized search trajectory to prevent overfitting during agent evolution
• 02Open edit space allowing modifications to prompts, tools, memory, and control flow
• 03Git-based candidate management for auditability and version control
• 04Multi-domain support including coding, workspace, and engineering design
• 05Automated critic and analyst components for screening and evaluating harness edits