Meta Hyperagents (DGM-H) Generalizes Self-Improvement Beyond Coding Tasks
Meta's Hyperagents (DGM-H) merge the task-agent and meta-agent into a single editable program, allowing the improvement mechanism itself to evolve — not just the task-solving code. During training runs, Hyperagents independently evolved persistent memory modules, performance tracking systems, and multi-stage evaluation pipelines without explicit human instruction to do so. Performance on paper-reviewing went from 0.0 to 0.710 accuracy; quadruped locomotion reward improved from 0.060 to 0.372, exceeding the human-designed baseline of 0.348. The system extends Sakana AI's Darwin-Gödel Machine, which improved SWE-bench from 20% to 50% but remained limited to coding tasks.
Why It Matters
Hyperagents achieving generalisation across domains — paper reviewing, robot control, and coding — suggests self-improving agents are no longer a coding-specific phenomenon. Independently evolving memory and evaluation infrastructure is the clearest evidence yet that agents can architect their own harness without human scaffolding.