Can AI Self‑Evolve? New Meta Research Redefines Agent Rules

The paper presents HyperAgents (DGM‑H), a novel agent architecture developed by Jenny Zhang during a Meta internship together with researchers from Meta AI, UBC, and NYU. Unlike previous self‑evolution systems such as the Darwin Gödel Machine (DGM), HyperAgents combine the task‑solving agent and the meta‑agent into a single editable program, allowing the system to modify not only its solutions but also the very mechanism that proposes improvements—a capability the authors call metacognitive self‑modification.

HyperAgents remove the assumption that task performance and self‑improvement are naturally aligned. In DGM, the instruction‑generation component was hand‑crafted and fixed, creating a bottleneck. HyperAgents instead make this component mutable, enabling the system to evolve its own improvement strategy.

Experimental results demonstrate that HyperAgents achieve comparable or better gains than DGM on coding tasks: on the Polyglot benchmark, performance rises from 0.140 to 0.340 on a 50‑task subset and from 0.084 to 0.267 on the full set. Crucially, the framework also succeeds in non‑coding domains. In paper‑review experiments, the initial agent scores 0.0 while HyperAgents reach 0.710; in robotics reward‑design, scores improve from 0.060 to 0.372. Transfer experiments to Olympiad‑level math grading show that, after 50 generations, the best generated agent attains an importance‑at‑50 (imp@50) of 0.630, far surpassing DGM‑custom baselines.

The authors highlight that these improvements stem from the system’s ability to grow its own infrastructure. During iterations, HyperAgents automatically generate performance trackers, persistent memory, evaluation analysis modules, and compute‑aware planning components. The persistent memory records cross‑generation metrics, identifies effective strategies, diagnoses over‑correction, and formulates subsequent improvement plans, turning isolated patches into a continuous optimization process.

Risk analysis is included: as agents gain the capacity to alter their own improvement loops, their evolution speed may outpace human auditing and understanding. Current experiments run in sandboxed environments with human oversight, but the authors warn that future open‑ended self‑improvement could pose safety challenges.

Overall, HyperAgents represent a shift from “who can build a stronger agent” to “who can build an agent that continually enhances its own improvement machinery,” potentially redefining competitive dynamics in AI research and industry.