Seven hours a day with a coding agent, and never a usage limit

The forums this week are full of people who hit a wall. New limits, shorter sessions, work stopping mid-task. I read twenty-five of those threads dated 15 and 16 September. I am on a flat-rate plan, I run the coding agent most of my waking day, and I have not hit a limit once since May.

That sounds like a boast, so I went and measured it instead of claiming it.

Engaged time is the union of five-minute activity windows across every session running at the time. If three sessions are live at 14:20, that minute counts once. Summing the sessions instead reported August at 24.4 hours a day, which is more hours than a day has, and I quoted that number before I caught it.

Measured from the session transcripts on 17 September 2026:

Those hours count only sessions in which I typed at least one prompt. A first pass counted every session and every prompt, and it turned out that 9,039 of 15,270 lines logged as prompts were machine-fired: a heartbeat poll that pinged an empty file 1,700 times in June, automatic continuations after a context reset, and compaction summaries. The instrument counted itself. A second pass found another layer: I had a habit of sending one prompt into several parallel sessions at once, up to eight copies of the same line in the same minute, and the log counted each copy. Distinct prompts from July to mid-September: 4,175 (391 in July, 2,398 in August, 1,386 in September); 6,246 counting the copies. The agent made 130,184 tool calls over the same span. Shell commands were 80.8 per cent of those calls. That is the shape of the work: reading files, running scripts, checking the output, trying again. Prose is a thin slice of it.

This is the part most people miss, and it took me 13,379 measured calls to see it.

Input was 91.4 per cent of what I spent. The cache read 3.62 billion tokens against about 13 million produced, a 278 times re-read ratio. Every tool call drags the whole conversation back through the model. A single session of mine ran 2,734 tool calls at 14.1 per turn, and each one paid for the context again.

So the lever is the call count and the size of what comes back, and both sit at the call site. Two habits do most of it. Batch independent probes into one call rather than six. Pre-slice tool output before it enters the transcript, with head, cut or a targeted grep, so a listing arrives at twenty lines instead of two thousand. Shell output was about 69 per cent of my context. Twenty-five uncut results at twenty thousand characters cost me 10 per cent of a window for three days.

A failed call costs twice, the attempt and the retry. A stable fact gets measured once and written down. I found 54 near-identical probes of the same machine setting in one sample, each one paid for in full.

The second reason is that most of the work never touched the expensive model at all. I route cheap work down cheap lanes: a second vendor's agent for bounded review, local models for text with no web in it, a job queue for anything batched, small subagents on the cheapest tier for parallel reading.

The queue line is the one that matters. 1,383 jobs in seventeen days, none of them billed against the allowance. The same seventeen days carried 47,497 tool calls and 82.5 million output tokens, which is the highest throughput per day I have recorded.

Add a caching proxy in front of repeated context, and one repository per prompt so the session never loads three codebases it does not need, and the picture is complete. The expensive model gets the judgement calls. Everything else goes somewhere cheaper.

My read of the complaint threads: the people hitting limits are running every kind of work inside one session on the expensive path. The routing is worth more than the model.

Three of them, and I would rather say them than have someone else find them.

The time figure counts the agent's activity, and I have been calling it my day. When I measure only the windows around my own prompts, the median is 4.1 hours. So roughly three hours a day the machine is working and I am somewhere else. Seven hours of agent time is real. Seven hours of my attention is a different claim and the data does not support it.

A cheap lane produces output I then have to check. Batch reader jobs I ran this month quoted 97 per cent of lines verbatim, put 27 per cent of quotes at the wrong line, and dropped 3 per cent entirely. One batch of 82 files returned template text and still printed a completion message. Cheap lanes are reliable on content and unreliable on citations, and the verification time is a cost that does not appear in any token count.

Volume and value came apart. Over the same months, the share of my messages that are corrections of the agent went from 15.5 per cent in July to 48.9 per cent in September. A count of how many items in my prompts actually get addressed came back at roughly one in three, and it is worst in the repositories where I load the most items into a single message. Some of that rise is my own standard moving: the work got more correct while I got harder to satisfy. The rest of it is real. Output went up about sevenfold. The number of things I would point at and call shipped, over any given fortnight, I can count on one hand.

Seven steps, in the order I would do them again.

1. Measure your own use before you argue about limits. Union the concurrent sessions. Summing them will hand you a number larger than the clock.

2. Cut calls, then cut context per call. Batch independent probes. Slice output at the call site.

3. Move bounded, checkable work to a cheaper agent or a local model. Review, batch reads, format conversion, anything with a known-good answer.

4. Put a queue between you and the work that can wait. Mine did 1,383 jobs last month off the allowance.

5. One repository per prompt.

6. Write the definition of done in the prompt, in your own words, before the first tool call. A shorter prompt makes the agent infer it from your rulebook instead, and my prompts under 200 characters produced twice the file sprawl of prompts over 600.

7. Check the cheap lane's citations, not its prose. That is where it fails.

Two numbers, both of which I have baselines for. The share of my messages that are corrections, at 48.9 per cent in September. The share of my items addressed, at about one in three. I am moving to a different agent on the 20th, and if neither number moves by ten points, then I changed the vendor and left the work alone.

The routing is what buys the headroom. It is also the reason I finished this month with four days of unused capacity while other people were locked out of theirs. What I do not yet have is evidence that the extra seven hours produced seven hours of things worth keeping. Those are two different questions, and only the first one is settled.