Furong Huang on Self-Improving AI Agents

Furong Huang posted on X that most AI agents adapt inside a single task by inspecting errors and switching tools yet lose that experience when the task ends. She identifies self-improving agents as the next frontier: systems that convert the consequences of today’s work into better methods tomorrow. Huang states the object of improvement is the whole agent—skills, workflows, action policies, evaluators, and sometimes parameters. She links to her Furong Lab blog post “Self-Improving Agents: Learning How to Work,” which examines how to test whether later performance actually improves on unseen tasks.