Meta's HyperAgents: When Agents Design Their Own Harness

HyperAgents, jointly announced by Meta and UBC, is not simply an agent that performs tasks, but a self-referential AI agent framework that modifies even the improvement mechanism itself.

The key point is that the agent is not tied to a fixed harness, and instead redesigns its own improvement loop as needed. In other words, it doesn't just change the task code — it also reworks the very structure that runs for better self-improvement.

The scope of application extends across multiple domains, including coding, paper review, robotics, and math grading. The notable point is that after repeating self-improvement across multiple domains, the agent showed behavior that increasingly and stably improved its performance.