Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents
A self-improving coding-agent system that modifies its own code, validates changes on benchmarks, and keeps a growing archive of diverse agents.
Can self-modifying agent archives continue producing transferable capability gains after benchmark feedback becomes sparse, noisy, or vulnerable to overfitting?
DGM is one of the clearest empirical steps toward recursive self-improvement: the system edits its own codebase and validates those edits against software engineering benchmarks.
Results
The reported run improved SWE-bench performance from 20.0% to 50.0% and Polyglot from 14.2% to 30.7%, while maintaining an archive of many generated agents instead of climbing a single hill.
ASI relevance
Open-ended archives matter. ASI progress may depend less on one perfect update and more on preserving diverse stepping stones, including changes that look locally odd but unlock later improvements.