Meta Hyperagents: Self-Referential Agents That Improve How They Improve — Metacognitive Self-Modification (arXiv 2603.19461)

arXiv·high signalXBlueskyLinkedInCopy link

Meta AI published Hyperagents, introducing DGM-Hyperagents that integrate a task agent and a meta agent into a single editable program, enabling metacognitive self-modification — the agent learns not just to perform better but to improve the mechanism that generates improvements. Tested across coding, paper review, robotics reward design, and Olympiad-level math grading, DGM-H outperforms baselines without self-improvement and prior self-improving systems. Code released at github.com/facebookresearch/HyperAgents. This is the strongest formalization yet of recursive self-improvement in agent systems.