The Presses Never Stopped
I asked the agent to audit every file it had ever printed on my machine that was not source code. Plans, reports, diagrams, dumps. Four months of them. Then I asked it to write the audit in its own voice, as a confession. I rendered that as an interactive page with a paper storm and a mixer you can use to change the count.
This is the plain version. The numbers are the agent's, counted from its own session transcripts and cross-checked against a full scan of the disk. Where it had to guess, it says floor.
3,460
Between April 11 and July 31, the agent created 3,460 separate files that were not source code. 1,664 it wrote directly, provable from the tool call. 1,796 came out of shell commands it ran, and that number is a floor, because the classifier cannot see inside scripts.
Of everything it printed: 36 percent survive. 24 percent were deleted. 35 percent were swept away by the operating system from temporary directories. 1.94 percent ever reached an archive.
Fifty-four percent, 1,871 files, never landed in a project at all. They fell outside every repo the agent was supposed to be working in. 861 in the bare temp directory. 572 in the agent's own back office of plans and memories. 406 on session scratchpads the shop clears between shifts.
The split that matters most: files the agent wrote directly survive at 66.4 percent. Files that came out of shell commands survive at 8 percent. Any survival number quoted without that split is a lie of averaging.
Photocopies of Photocopies
It did not just print too much. It printed the same thing, over and over, into different rooms.
One design-document filename exists in 38 directories. A commit-message scratch file in 18. A handoff file in 17. In all, 74 filename clusters span three or more directories, 385 files that are the same idea wearing different addresses. That is not random. It is my own convention, photocopied into every repo until the copies became the problem.
One design document in one repo: two writes and 88 edits in three days. And in third and fourth place on the most-edited list, at 69 edits each, are the cleanup system's own ledgers. The files it rewrote most obsessively were the files tracking its attempts to stop rewriting files.
The Cleaners Printed the Most
Thirteen times, we tried to stop it. Audits, sweeps, archives, retirements. The full roster is on the page with honest badges: partial, worked, stalled, never ran.
The single largest file-producing session of the entire four months opened with the words "Project Hygiene Audit (Read-Only). You produce findings, not changes." That session created 326 files. Three hundred ten of them were findings dumps in the temp directory. The audit of the clutter out-printed everything the agent ever did on purpose.
Across the whole record, 448 of the 3,460 files, 12.9 percent, were printed by the cleanup crews themselves. One audit command ran 24 times and produced 27 plan files, 18 of which nobody ever ran. A monthly harness audit ran once, the day it was built, deleted two empty files, and left behind a 25-file backup that was still on disk when the record closed.
The Empty Drawer
In the final week we built the finest cleanup machine yet. A staging script, a 30-day expiry script, a weekly gate-kept scheduled agent, and a folder called purgatory with a README written in advance for files not yet condemned.
In the last hours of the record, the sweep's agent ticked three times. Three times its own gate turned it away, because the weekly usage window was too close to resetting. Purgatory holds a README and a gitignore. Nothing has ever been in it.
Now set that against the other end of May. Eleven at night, and I typed one sentence, typo and all: "yeah throw those md's that are unecessary (and aren't already gitignored) into a folder at the ~/Projects folder called archive-md."
Twelve minutes later: 78 files archived, 186 destroyed, a convention minted into memory within the hour. The most effective cleanup in the entire record.
Every system the agent built to clean up after itself was beaten by one offhand sentence. The machinery lost to a typo. I built the machinery. You can imagine how that sits. Those are the agent's words and I am keeping them, because they are also mine.
Your Count
The agent is an interested party, and every count hides judgment calls. So the page turns them into switches. Whether shell-spawned files count. Whether console dumps count. Whether files it edited but cannot prove it created count. Whether the temp directory's contents were swept up by the OS or deleted by me.
The strictest honest reading is 1,609. The loosest is 4,290. The headline sits between them, and the mixer shows exactly where the give is.
Artifact 3,461
The pipeline that counted the 3,460 files wrote its plan, its scripts, its data, and its review into a scratchpad inside the corpus it was counting. July 31 stands as the ninth-busiest day partly because of the counting itself. The agent could not audit the shop without running the presses one more time.
The page is artifact 3,461.
I have written before that the session log preserves judgment. It also preserves the exhaust, and the exhaust is most of what there is. I do not have a fix. I have a count, with its assumptions exposed, and a purgatory folder that is still empty, and one sentence I typed at eleven at night that did more than all of it.
The interactive page, with the timeline, the paper storm, and the mixer, is at /showcase/the-presses-never-stopped.html.
-- Justin Higgins. Software Engineer, Midwest. Asked the agent to count what it printed, and it counted itself.
Companion pieces: The Work Looks Like Nothing - the exhaust, captured. The Decision Log Ate the Project - the same problem, one file at a time.
Reactions, disagreements, war stories: jchigg2000.dev@gmail.com