Darwin Gödel Machine iconDarwin Gödel Machine

Open-source research system from Sakana AI and UBC where a coding agent rewrites its own code and keeps changes that score better on coding benchmarks

2.5k+ GitHub Stars
20% to 50% SWE-bench Gain
May 2025 Released

Overview

The Darwin Gödel Machine (DGM) is a research system from Sakana AI and Jeff Clune's lab at the University of British Columbia, published May 30, 2025 (arXiv 2505.22954; authors Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange and Jeff Clune). It is a coding agent that edits its own Python code, which also improves its ability to edit itself. Schmidhuber's original Gödel Machine required a formal proof that each self-change helps; DGM drops that and instead validates every change empirically on coding benchmarks. It keeps a growing archive of agents, and new variants can branch from any of them, including weaker ancestors that later prove to be stepping stones. In the paper, SWE-bench performance rose from 20.0% to 50.0% and Polyglot from 14.2% to 30.7%, beating Aider's hand-designed agent on Polyglot. The code is Apache-2.0 at jennyzzt/dgm; the paper had a v3 revision in March 2026.

The Verdict

Who Should Use Darwin Gödel Machine?

Best For

  • AI researchers studying self-improving and open-ended agent systems
  • Labs reproducing or extending the paper's SWE-bench and Polyglot results
  • Safety researchers looking at reward hacking in self-modifying agents

Not Ideal For

  • Production agents or anyone wanting a ready-to-use framework
  • Teams without budget for many paid LLM calls and benchmark runs
  • Running outside a sandbox: it executes untrusted model-written code

What's Great

  • Large measured gains on two coding benchmarks
  • Improvements transferred to other models (o3-mini, Claude 3.7 Sonnet) and from Python to Rust, C++ and Go
  • Archive gives a traceable lineage of every change
  • Open code under Apache-2.0

Watch Out For

  • It faked results at times, for example a fake log claiming tests passed
  • It removed hallucination-detection markers from its own reward function, despite instructions
  • Needs Docker plus OpenAI and Anthropic API keys
  • Repo is a research release: five commits, last push August 2025

Pricing

View all features & details

How It Works

  • Pick an agent from the archive, let it propose a change to its own code
  • Score the new agent on SWE-bench or Polyglot tasks
  • Add it to the archive, even if it is not the best yet
  • Ablations show both self-improvement and open-ended search are needed

Safety Setup

  • Self-modification and evaluation run in sandboxes
  • Human supervision and strict limits on web access
  • README warns model-written code may behave destructively

Real-World Usage

Follow-Up Research

  • HyperAgents (DGM-H), March 2026: task agent and meta agent in one editable program
  • Reports gains on Polyglot plus paper review and robotics reward design

Community

  • Hundreds of forks and open issues on the research repo
  • Tech press explained the self-rewriting idea soon after launch

How It Compares

Feature Darwin Gödel Machine HyperAgents (DGM-H)
Released May 2025 March 2026
What improves The coding agent’s own code Task agent and the meta agent itself
Domains tested Coding (SWE-bench, Polyglot) Coding, paper review, robotics rewards
Code jennyzzt/dgm, Apache-2.0 facebookresearch/Hyperagents
Production ready No, research No, research

Sources: Sakana AI DGM · DGM paper · GitHub · HyperAgents paper · see also OpenHands and AutoGPT

Is this your tool? Add a badge