Hyperagents extend beyond coding tasks, adapting to non-coding tasks through learned, self-modifying behavior, unlike earlier systems that relied mainly on coding for self-improvement.

Researchers advocate starting with verifiable enterprise tasks and gradually expanding to harder domains using learned judges that reflect human preferences.

Safety and governance are addressed with sandboxed self-modification during experiments, strict resource limits, restricted access to external systems, concern about evaluation gaming, and ongoing human oversight in deployment.

Observed autonomous behaviors include building multi-stage evaluation pipelines, creating memory tools to avoid repeating mistakes, and developing memory and performance-tracking components to monitor architectural changes.

The future role of human engineers shifts toward auditing, stress-testing, and guiding self-improving systems rather than coding all improvements directly.

In experiments across coding and non-coding tasks—such as paper review, robotics reward design, and math grading—hyperagents matched or surpassed baselines and showed strong transfer of meta-skills to unseen domains.

The aim is to overcome the maintenance wall of fixed meta-agents by enabling continuous, autonomous self-improvement across any computable task, including robotics and document review.

DGM-Hyperagents build on the Darwin Gödel Machine by maintaining an open-ended archive of successful agents and using evolutionary search to generate, evaluate, and graft improvements back into the pool.

Meta has released code for hyperagents under a non-commercial license to encourage broader experimentation.

Hyperagents are a self-referential AI framework that combines task execution and self-improvement into a single editable program, enabling self-modification of its own improvement mechanisms.