Here, facts not feelings.
Across a selection of 5 of my projects writing code is on average 9.4% of total agent hours.
"Writing code" measures the time between the preceding logged event and emission of an edit-bearing response (apply_patch, including wrapped calls).
"Other model work" covers reasoning, code inspection, planning, and composing messages.
Parallel agents count separately, so these are agent-hours rather than elapsed clock hours. Interrupted turns are included through their last recorded event, making those contributions lower bounds.
I could break down "other model work" further into reasoning, code reads, planning, etc. because that would determine how much faster reasoning would affect the total agent hours, but I think my point stands, you could have all the code in 0.1ms but you'd still be waiting for tools and execution
