Keep the Limits, Delete the Roadmap

Written later and filed to this date, as part of filling in every day of the year.

You asked what to do about the documents. I know the ones you mean. Two days into giving an agent a repository of its own, you have a roadmap of agentic patterns, two fix logs, a self-restart design, a self-debugging design with separate implementation notes, a web-search design with the same, and a session summary whose first heading is “What You Asked For”, with your sentence quoted and a checkmark under it. Mine came to about twenty-five hundred lines in forty-eight hours, for a library of about two thousand lines of code. One thing only: delete them, and keep the limits.

By delete I mean a commit, not an archive. Mine was called init, fourteen commits in, 2,770 lines out and 2,264 in, and the code it removed was zero lines. What went in was structure for the second attempt, some tests, and a README with two sections that had not existed before. “What works today”: three bullets, each a thing you could run. “Known limits”: three more. Requires the command line tool installed and logged in. Improvement quality depends on the tests and linters available. Not a general multi-model graph. Those six bullets were the first sentences in the repository that described it as it was.

Here is why the long documents are worth less than they look. They were written by the thing they describe, about what it intends to become, and intentions are the cheapest output a model has; regenerate them any time by asking. The limits are the opposite. They are what the machine found out by failing, and nobody will write them for you later, because by then they will be obvious to everyone except the stranger reading the log.

One warning about the replacement. The same commit that gave mine its limits also gave it a reliability table: three rows, date, metric, value, how measured, under a line that says to update them regularly from real runs. The dates are placeholders. The values are empty. No commit since has touched that table. A table with the right columns and no rows is a promise addressed to nobody, and it will sit in your README looking like rigour. Fill the first row before you commit it, or leave it out.

The last thing to know is that the deletion will not slow the agent down. Mine committed three more times that evening, after init, each a fix it called P0, under the same subject as the three before. What changed was not its pace. It was what the next reader would find beside the code: not what the machine meant to become, but what it needed in order to run.


← all writing