According to God of Prompt on X, Meta demonstrated DGM-Hyperagents, a system where the improvement mechanism can rewrite itself, removing the long-standing architectural bottleneck in self-improving AI. As reported by the posted thread, prior designs like DGM, ADAS, and Gödel Machine variants fixed the meta agent by hand, limiting open-ended optimization; DGM-Hyperagents merges task and meta agents into one editable program, enabling metacognitive self-modification. According to the same source, the system autonomously built persistent memory, performance tracking, and compute-aware planning to accelerate improvement. The thread reports a transfer test where a hyperagent trained on paper review and robotics achieved imp@50 of 0.630 when dropped into Olympiad-level math without prior exposure, compared with 0.000 for both original DGM transfer agents and an untrained initial agent. According to the ablation cited in the thread, removing metacognitive self-modification or open-ended exploration reduces paper-review performance to 0.0, while the full system reaches 0.710, indicating both components are necessary. As reported by the thread, Meta sandboxed all experiments with human oversight and kept parent selection fixed outside the system’s control, suggesting a constrained safety setup. If validated by Meta’s publication, the business implications include faster R&D loops for enterprise automation, adaptive agent platforms that self-architect memory and tooling, and cross-domain transfer focused on learning-to-improve rather than task knowledge, creating opportunities in AI Ops, robotics, and developer tooling.