Fable Decides, Opus and Sonnet Do the Work: How I Route Claude Code Subagents (www.practicalsystems.io)

A recent discussion on Claude Code's architecture reveals significant updates in effectively managing AI agent roles within workflows, particularly highlighting the optimization of using models like Fable, Opus, and Sonnet. The author shares insights from lessons learned through a previous inefficient use of Fable, where an entire session was monopolized by one workflow, resulting in excessive costs and resource consumption. By restructuring the roles of these agents, Fable is now designated primarily for high-level decision-making and oversight, while Opus and Sonnet handle the execution of specific tasks. This delegation minimizes the expensive context re-reading Fable was previously subjected to, thus conserving resources and improving efficiency. The technical implications are profound; the new structure allows for a more streamlined workflow where subagents operate independently without burdening Fable’s context with raw data, thereby reducing the overall token consumption significantly. Each agent now clearly delineates its scope and tasks: Opus for large, complex tasks, Sonnet for precise coding once decisions are made, and Haiku for data retrieval. The addition of a built-in advisory system further refines the process by ensuring consultation flows appropriately between agent tiers while maintaining a focus on cost-effectiveness and resource management. This reorganization exemplifies the critical importance of effective agent collaboration in AI workflows, setting a precedent for future optimization strategies.