HyperAgents extends DGM with editable self-improvement loops

This is a smart step past fixed-loop self-improvement: the real bottleneck is no longer just getting an agent to improve, but letting the improvement machinery itself evolve.

• –The paper removes the handcrafted meta-level mechanism that classic DGM-style systems depend on, which is the biggest conceptual leap here.
• –Reported gains across diverse domains matter more than a single benchmark bump because they suggest the loop can transfer beyond coding-specific setups.
• –The open-source repo makes the idea hackable for others, with separate task-agent and meta-agent code paths plus scripts for running the loop.
• –The hardest follow-up problem is control: once an agent can rewrite its own improvement process, evaluation, rollback, and safety become first-class design constraints.