Definition
Plain language
An earlier system where an AI rewrites the scaffolding around itself to get better over time.
As stated in the literature
A self-improving agent framework in which an AI edits its own surrounding code (the harness) while the base model stays frozen; cited as prior work in the scaffold-only branch that SIA unifies with weight-update self-improvement.
Why it matters: It shows a model can improve its own performance by upgrading its surroundings, even while its core stays untouched.
For example, instead of changing what the model knows, this kind of system rewrites the tools and instructions wrapped around it to make it more effective.
Heard on the show
“Think Darwin Gödel Machine, Meta-Harness, the AI Scientist line of work.”Episode 088 — Two Levers for Self-Improving AI: When Rewriting Code Isn't Enough