Meta's Hyperagents Don't Just Improve at Tasks — They Improve at Improving
Researchers at Meta and the University of British Columbia have built 'hyperagents' that can rewrite both the task-solving part of their code and the mechanism they use to improve. Unlike prior self-improving AI, the optimization loop itself becomes subject to optimization — breaking through the ceiling that has limited recursive self-improvement since the concept was first formalized.
D.O.T.S AI Newsroom
AI News Desk
Self-improving AI systems have always run into the same wall: the mechanism that drives improvement is written by humans and never changes. The agent can get arbitrarily good at the task it was designed for, but it can never escape the constraints of the fixed loop that governs how it improves. A research team from Meta, the University of British Columbia, and several partner institutions has published a method they believe breaks through that ceiling.
The approach, which the team calls DGM-Hyperagents (DGM-H), builds on the Darwin Gödel Machine — a prior method that showed a coding agent could improve itself through repeated self-modification. The agent generates variants of its own code, tests them, and archives successful versions as stepping stones for further refinement. DGM-H extends this by making the improvement mechanism itself part of what gets optimized. Both the task-solving code and the code that modifies the agent live in the same editable program — so when the agent rewrites itself, it can rewrite the meta-level too.
Why Prior Approaches Couldn't Generalize
The original DGM worked well for coding tasks, where being a better programmer naturally makes you better at writing self-modifications. That link breaks down in other domains. An agent that gets better at evaluating scientific papers doesn't automatically get better at rewriting its own code. The team found DGM hit near-zero performance on non-programming tasks without manual tweaking.
DGM-H sidesteps this by making the improvement mechanism itself improvable — independently of what task the agent is doing. The team tested this across four task areas, demonstrating that the self-accelerating loop generalizes beyond coding.
What It Could Mean
The implications for AI capability trajectories are significant, if still early. Systems that improve at improving — rather than just improving — could in theory reach capability thresholds faster than scaling compute or data alone would allow. The research is published and peer-reviewed; production deployment is a different question. But Meta's willingness to publish on recursive self-improvement mechanisms signals that the lab views this class of research as publicly defensible, which itself tells you something about where the frontier is moving.
