A self-improving coding agent from Jürgen Schmidhuber's lab that approximates the optimal Gödel machine via clade-metaproductivity (CMP) search — scoring self-modifications by the downstream productivity of their descendant lineage rather than immediate benchmark gains. The result reaches human-engineered-agent-level performance on SWE-bench.

The 401★ metauto-ai/HGM release drew clean-room reimplementations and press coverage. It is distinct from the separately tracked Darwin Gödel Machine (Sakana AI / UBC), which pursues a related but independent self-improvement line.

Paper

agentsresearchcodeself-improvement