ASI·Research
ASI-LIB-014 technical code available

Hyperagents

Jenny Zhang, Bingchen Zhao, Wannan Yang, Jakob Foerster, Jeff Clune, Minqi Jiang, Sam Devlin, Tatiana Shavrina

Hyperagents system diagram comparing DGM and Hyperagents
Figure via ar5iv rendering of arXiv:2603.19461

Hyperagents generalize the DGM idea beyond coding tasks. The central move is to put the task agent and the meta agent inside a single editable program so the system can improve both what it does and how it improves.

Why it matters

DGM works especially well when coding skill and self-modification skill align. Hyperagents attack the harder case: any computable task where the path from task performance to better self-improvement is not obvious.

Code

The public repository includes domain code, task/meta-agent implementations, and a generation loop. It also carries the right warning: generated code should be executed in controlled environments.