Discussion: Recursive Self-Improvement Will Start in the Harness Layer

sudoraohacker · x · 2026-07-31

Yoonho Lee argues that Recursive Self-Improvement (RSI) will likely start in the harness layer rather than pre/post-training code. The harness is the cheapest place for an agent to test hypotheses and fork its own behavior at scale.

He notes that a lot of information must live outside the model (because it's private, too specific, or changes too fast), meaning we will always need a "fat harness" to handle irreducible task complexity.

Original post →

More from AGI Musings

AGI Musings channel →