Hyperagents fuse task-solving and self-modification into one program

A hyperagent integrates a task agent and a meta agent that modifies both itself and the task agent into one editable program, so the procedure that generates improvements is itself open to improvement. Meta researchers (Jenny Zhang, Jeff Clune and colleagues) instantiate it as DGM-Hyperagents, extending the Darwin Godel Machine beyond coding, where task skill and self-modification skill happen to coincide, to any computable task; across domains it beats non-self-improving baselines and prior self-improving systems, and meta-level gains such as persistent memory and performance tracking transfer across domains and accumulate across runs (arXiv). It generalises ALMA’s meta-learning of memory design to the whole improvement mechanism, arrives a day after MiniMax’s self-evolving M2.7, and is the theoretical form of what June’s self-rewriting harness later showed in the wild; the recursive-self-improvement theme picks the mechanism up again in a research loop that writes the strategies for its own outer loop.