During the Christmas break, a developer explored the implementation of a player/coach workflow using Kiro CLI subagents, inspired by agentic tools and reinforcement learning. This approach involves generating output from a 'player' agent, which is then assessed and reviewed by a 'coach' agent, creating a feedback loop to improve outcomes. The developer created three custom agents: the orchestrator, player, and coach, each with tailored system prompts and contexts.

The orchestrator initiates tasks, the player executes them, and the coach reviews and provides feedback. This setup was designed to enhance the quality of output by leveraging the strengths of both agents. The developer shared their experience, including challenges like 'hanging' services, file creation issues, and cost considerations.

They also highlighted the importance of specific evaluation criteria and steering documents for better performance. The post invites others to experiment with the provided code and resources, encouraging community engagement and exploration of this innovative workflow. This player/coach model represents a promising direction in agentic AI, potentially leading to more refined and efficient outputs in various applications.