Skip to main content
There is no scheduler screen and no cron tab. A job that runs on its own is a program in your harness — a loop that wakes, decides whether today’s work is due, does it, and sleeps again — and like everything else here, you get it by asking for it. What you end up with is a directory on your machine that gains a file every morning, with nobody watching.

​
What you say

Three things in that sentence are doing the work, and they are the three worth naming in any job like this: when (every morning), what it leaves behind (a file on the machine), and that it runs without you — which is what makes it a program rather than something you ask for each day. The fourth clause is bookkeeping, and it matters more than it looks: take the old one out. A harness accumulates. Say what should stop, not only what should start.

​
What came back, and the surprise in it

The agent read the old programs/nightly before replacing it — and read the file that program had actually produced that morning. The briefing said the day came to 163.50 across three expenses. The ledger it was written from has four rows:
The refund was not in the total, and it was not mentioned. The old program had an agent read the CSV and write the prose, so the day’s arithmetic was something a model did in its head, once, unwitnessed, into a file nobody would check against the source. The replacement computes every number in plain Python and has the tests to say so. No agent at all. Nothing in the request asked for that; the agent worked out that this job did not need a model and said why.
The maths lives in brief.py, beside main.py in the program’s own directory, and it has a test file next to it:
Its own docstring puts it plainly:
Everything numeric happens here, in plain code, where a wrong sum is a bug the tests catch — not a sentence a model once wrote.
A program directory may hold anything. main.py is the only file the platform opens; files beside it are imported by plain name. See Programs.
This is the judgement worth borrowing: a model is for the part that needs judgement. Summarising a week of email needs one. Adding up four numbers does not, and a model that does it anyway will be right most mornings — which is worse than being wrong every morning, because you stop checking.

​
Publishing and starting it

Publishing is a push, and the tree it pushes into holds programs, tools, evals and root documents — nothing else. This push was refused once for a root .gitignore, so the ignore rules moved inside the program’s own directory, where programs may be any shape they like.
A Splox chat where the agent publishes version 9 and starts the program

The refusal, the fix, and the version the push created

Publishing and running are two different things. A published program is a version of your harness; a running program is a process on your machine, started once and staying up between chats:
The agent starts it for you when you ask it to. It then started it a second time to show that doing so is harmless — the second instance sees the first and exits.

​
Showing that it works

The proof is not that a process is listed. It is that the file changes when the world does. The agent appended a row to the ledger, waited for a tick, and read the briefing back:
Then it took the row back out, waited for another tick, and the file returned to the truth:
Both numbers the old program conflated are now in the sentence, and which one is net is stated rather than implied. The day’s file also stays current as the ledger grows, instead of being written once at seven and going stale — another thing nobody asked for, and the right call.
“Every morning” is the machine’s clock, which is UTC. The demonstration above is ticks a minute apart, not a wait until tomorrow — if your morning is a particular morning somewhere, say so in the message, because nothing else in the system knows your timezone.

​
Checking on it later

Ask. Your agent is on the same machine as the job, so “is the briefing program still running, and what did it write today?” is a question it answers by looking — the process list, /tmp/spend-briefing.log, and the file itself. To stop it, say so. There is no jobs screen in the app. This is not an omission being apologised for: the loop, the schedule and the log are files in your harness, so the thing that knows about them is the agent that wrote them.

​
Next

A bot that answers from outside

The same shape with a different trigger: a message from Telegram instead of the clock.

Programs

What a program is, what handle(msg) decides, and how a program calls the agents it declares.