Darwin Gödel Machine
#28 in Agent FrameworksSakana Ai · 3× · last seen Jun 30, 2026
The Darwin Gödel Machine (DGM) is a research system for self-improving Coding Agents developed by Sakana AI in collaboration with the University of British Columbia and the Vector Institute. An Agent iteratively modifies its own Python source code and empirically validates each change against Coding Benchmarks such as SWE-bench and Polyglot, rather than formally proving improvements as in the theoretical Gödel Machine concept. The system maintains a growing archive of different Agent variants, from which new, improved descendants are generated through open exploration. The complete code is Open Source and available on GitHub; all self-modifications and evaluations run in sandboxed environments under human supervision.
Features
| Autonomie-Grad | Hoch: Agent modifiziert eigenständig seinen Code, unter menschlicher Aufsicht in Sandbox mit begrenztem Web-Zugriff |
| Lizenz | Code: Apache License 2.0 (GitHub); Paper: CC BY 4.0 (arXiv) |
| Modalitäten | Code/Text (Coding-Agent auf Basis eines eingefrorenen Foundation Models) |
| Plattform | Python-Codebasis, lokal/GCP-VM ausführbar, Sandbox-Docker-Umgebung |
| Release-Datum | 29. Mai 2025 (arXiv-Preprint & Sakana-AI-Blog-Launch) |
| Tool-/MCP-Integrationen | Werkzeugnutzung wie Bash-Kommandos und Dateibearbeitung; selbst entwickelte Tools (Patch-Validierung, Datei-Viewer, Editier-Werkzeuge) |