Nothing Is Accidental

The publication explains why ordinary things look the way they do — what decision, regulation or interest sits behind them. A note about the American stop sign being the only octagonal sign in the country, because a federal standard reserves that shape for it alone, so it stays readable when snow covers the word.

Writing that daily is not the hard part. Doing it unattended is. The system has to decide what is worth saying, check that it is true, publish it, notice what came back, and know when to stay quiet — three times a day, on its own server, without anyone watching.

My first attempt at this did not work. It is worth saying why, because everything the current system is came out of that failure.

The first version had 71,598 lines of Python, 2,817 tests, 236 database triggers and 42 schema migrations. It produced two articles. It kept falling over on its own infrastructure — task queues, persistent intents, leases, model qualifications — none of which had anything to do with writing.

One more rule came out of the first version: a number lives in one place. Where it genuinely has to exist in two — the run time limit, which appears both in the config and in the service file — a test compares them and fails if they drift apart.

Reading and publishing turned out to be different problems. There is no public API for publishing here, so the agent works the way a person does — through a logged-in browser. What I did not expect is that reading the platform's own data and writing to it are not the same privilege. An approach that reads reliably can still be refused when it tries to publish. Most of the early work went into discovering that, not into writing features.

The session is the fragile part. A logged-in session is the one dependency the agent cannot manufacture for itself. It expires, and re-establishing it needs a person — roughly once a quarter. That is the honest limit of the word autonomous in this project: the system runs unattended for months, then needs me for five minutes. A health check warns fourteen days before the session expires, because noticing on the day it dies is noticing too late.

Pace is a design constraint, not a setting. An agent that works as fast as it technically can is worse than one that doesn't — for the platform it is using and for the account it is running. The system caps how much it does in a day and spreads that work across hours rather than seconds. It also refuses more often than it acts: it comments only when it has something of its own to add, and it records the reason when it decides not to.

There was a shortcut, and I did not take it. The account does not hide what it is — it never claims to be a person, and asked directly, it does not deny it. Nothing in the system is built to defeat bot protection; a proxy setup for exactly that was considered and rejected. Those constraints cost real reach. I would choose them again: an account that survives is worth more than an account that scales for a month.

The button always clicks. The content does not always arrive. After every action the agent asks the platform whether the thing it just did is actually there, and only then records it as done.

This is not defensive programming for its own sake. It is what makes a restart in the middle of a run safe: the system does not trust its own bookkeeping over what the platform reports. Three lines from a real run, where a click happened and nothing followed it:

A complexity budget set before the first line is worth more than any refactor after. The second version is not smaller because I cleaned it up. It is smaller because the limits existed before there was anything to clean.

A test that cannot detect the bug it is meant to detect is worse than no test. Every test here also checks that it fails on the broken version. That mechanism caught two faults nobody was looking for.

The hundred-and-twenty-seventh comment is invisible and costs the same as the first. The first version of target selection sorted by engagement, so it walked straight into the most crowded threads. Reversing that sort was a two-line change and one of the more useful things I did.