That holds up for code generation (where tokens fly by), but not for tool use. The agent often stalls between tool calls, and those are exactly the moments I need to see what it's planning, not just stare at a blank screen