Fifty-nine per cent of the log was the tool talking to itself

I have been running a coding agent for most of every working day since May. In September I wanted one number out of it: how much of those five months went into the agent itself, and how much went into anything I actually sell.

So I had the transcripts read. Sixteen reader agents, one time window each, told to read every line rather than sample, and to mark any figure they had estimated rather than counted. A seventeenth landed mid-pass and got folded in. Twelve more readings covered the venture repositories over the same months. Every number below comes from those readings.

From 1 July to 17 September the logs hold 15,270 lines filed as prompts from me. Of those, 6,231 were typed by me, and only 4,175 were distinct, because I had a habit of sending one prompt into several parallel sessions at once. The other 9,039 were the machinery: 59.2 per cent of the log before the copies are even removed. Taken raw, the log overstates my typed input by 2.45 times. May and June contain zero human prompts at all.

The readers put names to 4,220 of the machine lines.

The heartbeat was my idea. A poll fired every 20 to 30 minutes, opened a task file, found whatever was in it, and replied that it was fine. The file stayed empty the entire time. One reader put the idle polling at about 2,500 billed turns across six weeks for zero output. Counting the May window as well, 3,077 logged turns were a gateway talking to itself. Nobody noticed the 3.5 days where it was not even logged in, because nothing downstream of it existed to break.

Automatic continuation after a context reset is the second machine class: 252 in a single eight-day window, and the same verbatim line returning later under other names. Fan-out is the third. One question typed once and sent into up to eight parallel sessions logs as eight prompts. In one July window, 229 logged blocks hold 29 real turns of mine, a ratio of 7.9 to one. In another, 590 logged pairs reduce to 33 distinct texts, 21 of which I composed.

Run that back over July and my headline of 1,017 prompts becomes about 366 real turns across the windows anyone read. Every prompt count I have ever quoted, including in something I published two weeks ago, carried this freight.

Strip the machinery and the second finding is worse than the first.

In the sessions outside my venture repositories, the agent's own tooling and rules and self-measurement are the work. In one 25-day window from mid-August, that category is 61 per cent of 720 real asks. Measurement of the agent alone is 221 of them, and 160 of those are tests of a single hook to see whether it would say ok. Rules and punishment is 139 in the same window. Venture work is 95, or 13 per cent. In an earlier week, tooling was 38 per cent of my records and self-measurement 31 per cent, so about seven in ten.

The trend is the part I would like back. Agent-directed work was 3 per cent of real turns in mid-July. It was roughly a quarter by early August. It was 61 per cent across that 25-day window.

What it displaced is specific. On one August afternoon I spent 90 minutes closing out an authorisation design while outreach sends stood at zero. On another day the log carries my own line, "not a single connection request or DM went out", timestamped the same day I spent the evening on session-identity plumbing. Two scheduled jobs failed for six and seven days running before anyone noticed, and the noticing became the deliverable. Across all twelve venture windows in the same five months: no sale, no placement, no funding, no revenue.

Some of it survived contact.

The point system belongs in the right-hand column and deserves its own paragraph, because I asked for it. I wanted a penalty severe enough to change the agent's behaviour. Ten days later its own instruments said it never tracked me: the correlation between the score and how often I pushed back was +0.02. The two arms of its own experiment came out at 42.7 and 38.2 per cent push-back with mean deductions of zero in both. Roughly 1,440 of 2,000 deducted points were one class, and the first event in that class was my eleven typed copies of one line being counted as me repeating myself. The score sat between 7 and 27 on six of nine days, which measures the floor. It is paused now.

The worst single item is quieter. A gate that refuses my close-outs sits on disk, is described in my rulebook as active, and was never wired in. Two readers found it independently, a month apart, by checking whether the file was actually loaded. The rule was real to me and had never once fired.

1. One repository per prompt. Where asks crossed repositories, the tool audit won the day: venture asks lost the same-day competition on 3 of 5 days in one August window. 2. Never fan a prompt into parallel sessions. Eight copies gave me 229 log lines for 29 real turns, and four sessions on one question produced four answers to reconcile by hand. 3. No scheduled job without a named consumer and a failure alarm. Two failed for 6 and 7 days before anyone noticed, and the noticing became the deliverable. 4. Write the definition of done before the first call. The worst ratios sit under "fix everything" and "implement all", which closed with "no code moved". 5. Grade on sent, deployed, sold. Close-outs say DONE; the outcome columns say 0 sales, 0 placements, 0 funding across 12 venture windows. 6. Cap agent-directed work at one day a week and hold the cap. It ran at 61 per cent of a 25-day window. 7. Correct the agent twice on one thing and stop the session. "i told u 3-4 times", "publish; well relitigate" 6 times, the basket widget 6 times, the decisions table 3 times: a repeat says the lane is dead. 8. Ban the measurement-instead-of-fix trade by name. If the ask was a fix, a measurement is a refusal, and at least nine days in this record went that way. 9. Kill any rule whose enforcement you have not watched fire. A gate sat unwired for over a month while the rules file claimed it was blocking my close-outs. 10. Read the machine share before quoting any usage number. 59 per cent of the log was the tool talking to itself, and every prompts figure I have used until now carried that freight.

An instrument that logs its own activity next to yours will flatter you, and it will flatter you in the direction of effort. Mine told me I was prolific in June, when what it was recording was a poll reading an empty file every half hour.

The fix is cheap. Count what the machine fired. Count what you typed. Publish the second number. Then check, once, that the rules you believe are protecting you have ever actually stopped anything.