AI agents can now learn how to improve themselves — instead of using rules humans wrote

Researchers built a system where AI language agents automatically discover the best way to fix their own mistakes during use, rather than relying on improvement strategies that humans hand-coded in advance. In practice, this means an AI chatbot or web agent could learn its own correction method from experience, and that method would work on new tasks it has never seen before.

For years, every improvement to language agents came from engineers manually designing the rules for how they should fix errors. This work shows those hand-written rules are bottlenecks — the system finds better correction strategies on its own by testing thousands of approaches across different task environments. What changes is that instead of a team of engineers deciding how an agent should learn, the agent discovers its own learning method, which then transfers to new problems. This removes a layer of human guesswork from the deployment cycle.