I am experimenting with involving subagents more in my llm coding agent workflow. I discovered that subagents make it more difficult to diagnose token usage in Claude Code11 Technically the same thing is true of Pi, but I haven’t had the same cost issues in Pi yet. I’ll write more about Pi, Emacs Agent Shell, open weight models, and privacy-friendly inference providers some other time., because the regular /context command only breaks down context usage for the main conversation, not for the subagents.