This framework enables AI agents to self-improve on unfamiliar tasks without changing model weights. It proposes task-specific modifications—such as adjusting system prompts, memory, skills, sub-agents, and plugins—then runs real trials to compare results, and finally captures the lessons learned for future tasks.