The Sequence Knowledge - Issue 945: Learning RSI: Agents that Rewrite their Own Scaffolding

Some of the most useful software work looks boring. Better file viewing. A patch check before a fix gets submitted. Trying several answers, then ranking them instead of grabbing the first decent one. Keeping a log of what was tried, and why it failed.

That list sounds like a senior engineer’s onboarding notes. It wasn’t. It came from a coding agent writing into its own codebase, unprompted, over roughly eighty iterations while nobody was watching. The agent was the Darwin Gödel Machine from Sakana and Jeff Clune’s lab.

The result was not subtle. On SWE-bench, its score moved from 20 to 50 percent. On Polyglot, it went from 14 to 31 percent. Those numbers are the story here as much as the code changes themselves.

The bigger point is that the machine did not behave like a sci-fi robot chasing some grand abstract loop. It behaved like a mechanic. It improved the tools around itself, step by step, in ways a human engineer would recognize immediately. That is less dramatic than the old recursive self-improvement fantasies, and more interesting too.

This is why the source keeps coming back to that checklist feeling. The agent’s own scaffolding became the work product. And that makes the whole thing feel less like magic and more like a very determined engineer that never got tired.