Lilian Weng outlines harness engineering for RSI

Orchestration, not raw model scaling, is the true bottleneck and frontier of agentic self-improvement, transforming software engineering into the ultimate reinforcement learning playground.

• –Harness engineering allows agents to self-improve dynamically without expensive weight fine-tuning or retraining.
• –Using the file system as persistent memory enables agents to manage long-horizon tasks and recover state after system interruptions.
• –Meta-optimization systems like Meta-Harness and AFlow demonstrate that LLMs can discover complex strategies, such as genetic algorithms, when given direct control over their own execution code.
• –The lack of fast, objective, and robust evaluators remains the primary obstacle to scaling recursive self-improvement outside of sandboxed coding and math tasks.