I run several coding agents in parallel, mostly Claude Code and Kimi, sometimes Cursor and Copilot. I felt slow, so I exported every message I sent them over 3 weeks and had each one read and labeled by what it was for. I did not use keyword search. I tried that first and it gave wrong numbers.
After removing automated traffic, 2,116 messages were really mine, about 96 a day.
What my messages were for:
55% real work: new tasks, questions, decisions 13% correcting the agent: wrong thing, drifted, changed the model or scope without asking 9.5% "what's the progress?" 6% carrying info by hand from one agent or chat to another 4% "go", "yes", "continue" because the agent stopped to wait 4% "explain that in simple English" 3% repeating a rule I had already given 5% other: slash commands, fragments
So 40% of my typing is overhead. Some details:
I asked for progress 200 times. More than half came in bursts within the same hour, while I sat waiting on long runs that finish silently.
The biggest single cause of corrections was UI work, fixed one screenshot at a time.
The same rules came back on different days and with different agents: "don't waste tokens", "ask before changing the model", "don't release until it's verified". A rule I give one agent never reaches the others.
On my worst day I relayed 33 times. Two agents were on the same experiment and I was pasting each one's report to the other.
Token anxiety is constant. Agents start big runs without saying what they will cost.
One surprise: more than half of the "user messages" in my logs were not me at all. They were scripts and test harnesses running under my normal config.
My questions:
Do you see the same problems, or is your overhead somewhere else?
Which one costs you the most?
Has anyone actually fixed the progress polling or the rules-don't-carry-over problem, especially across different tools?
Top comments (1)
The 6 percent spent carrying information by hand between agents is the line I would try to kill first. Everything else on that list is at least the work. That one is you being the integration layer.
The part I would keep is that you threw the keyword pass away. I have done the same on my own app's data and the keyword version always looks cleaner and is always wrong, because it counts phrasing instead of intent.