Read back from my own session logs over two and a half months. Token counts and tool calls only, no content of any kind. The short version: I mostly direct the model, and it does the rest.
Daily output tokens. The first six weeks are faint because I was still at my previous job and only opened it some evenings. In late April that changed, and it became how I work.
A model you ask a question is a chatbot. A model you hand a goal and a set of tools is something else. Four numbers that separate the two.
When one agent isn't enough I fan out: a workflow splits a job across many sub-agents that run at once, and those agents spawn their own. The work is mostly review.
Not tokens for their own sake. Real edits to real files, across the stack and across research, infra, and the web.
When a new model ships, I move to it. First time each model appears in my logs, usually within a day or two of release.
The tool-call breakdown across every session. Mostly running a shell, editing files, searching, and spawning other agents.
The same way of working spans CAD generation, financial knowledge graphs, robotics perception, and research systems. Names withheld here; sized by output.
A few of the things the numbers above turned into. Public where it can be.
The plain totals, for the record.