Process where an agent progressively improves its capabilities by refining task specifications based on past execution traces.