The first time I saw a scheduled task, I ignored it.
ChatGPT shipped the feature in beta on 14 January 2025. I looked at it and felt nothing. It did not make sense to me and I did not use it.
That was nineteen months ago. Today it is the first thing I try to get someone to leave with — every meeting at Charleston AI, every education class. Not a strategy. Not a prompt library. One scheduled task that runs daily and produces something.
The feature did not change that much. What changed was what was on the other end of it.
You cannot schedule an adviser
The reason a scheduled task felt like nothing in January 2025 is that ChatGPT was a chat product, and a chat product is a consultant.
It tells me how to do things. It is genuinely good at that. But a recurring appointment with an adviser who has no hands is a strange thing to want. What would it do at five in the morning? Compose advice into an empty room?
ChatGPT Work and Claude Cowork are not in advice mode. They are in work mode. They have tools, files, connectors, permissions, somewhere to put the output. They do not tell me how the report should be built. They build it.
Scheduling only becomes valuable at the moment the thing on the other end stops advising and starts working.
That is the whole explanation for why a feature I ignored became the thing I teach first.
Where a scheduled task is born
Every scheduled task I have starts the same way. I notice that I am doing something.
Not a pattern noticing. Me noticing. There is a piece of work I am consciously attending to, week after week, and at some point I see it clearly enough to realise I do not have to be the one attending to it.
Then I hand it over, and I do it by talking. It is the closest thing in the federation to speaking to a person. Here is what I want you to do. Do this work every morning at five.
I try to stay high level. I describe the outcome rather than the keystrokes. And I schedule deliberately around cost — I keep away from expensive hours, because at this volume the hour you pick is a real number.
I noticed I was doing it.
I decided I no longer wanted to.
I said so once.
That is a scheduled task. It is the most ordinary thing in this ledger, and it is where almost everyone should start.
Why I keep them narrow
Nearly all of mine are narrowly defined, on purpose. That is not a limit on what they could do. It is a decision about what I want to be able to control.
Once you have thousands of these, scope becomes the thing that determines whether you can still steer. If a task carries too much, I cannot pause it without losing five other things I did not intend to pause. I cannot rewrite it without a ripple travelling through the federation.
Narrow scope means every task has its own switch.
I can stop one. I can replace one. I can let one run untouched for a year. None of that disturbs the rest of the colony, because I built them so that it wouldn’t.
Do not schedule what you have not settled
This is the mistake I made, and it is the one I would most like to save somebody else from.
In the beginning I scheduled things too early.
Think about the bank. You would not set up an auto-draft for a bill you intend to log in and adjust every month — take a little off this month, change the amount, hold it back until the fifteenth. If you are going to do that, the auto-draft is not saving you anything. It has become a recurring chore with extra steps.
An auto-draft is for the payment you have settled. You know the payee. You know the amount. You have stopped having opinions about it. So you say it once and then you never think about it again.
My early scheduled tasks were auto-drafts on bills I was still arguing about.
I had not finished deciding how I wanted the work done. So I scheduled it, and then I went in and adjusted it, and adjusted it again. What I had built was not a delegation. It was a standing appointment with myself.
The test is not whether the work repeats. The test is whether you have stopped changing your mind about how it is done.
It is the same error as correcting a prediction machine twenty times in a row, committed on a clock instead of in a conversation. If a method matters, finish it first — then hand it over as a finished thing.
Which gives me a diagnostic I now use constantly:
If you find yourself editing a scheduled task,
you scheduled it too early.
Go and do the work by hand a few more times. Not because the machine cannot do it, but because you do not yet know what you are asking for. When the method stops moving, schedule it — and then leave it alone.
I did not have this discipline at the start. I have it now, and it is most of the difference between nine thousand tasks that run and nine thousand tasks that need me.
The three, finally distinguished
These three categories are the ones people conflate, and it is worth being exact, because they behave differently and they fail differently.
| Scheduled task | Absorbed action | Sovereign act | |
|---|---|---|---|
| Whose intention | Mine, stated out loud | Nobody’s — a pattern requires it | The entity’s, and not mine |
| Who chooses the method | I do | The pattern does | The entity does |
| Where it is born | In my attention | Beneath everyone’s attention | In the entity’s judgement |
| How I find out | I set it up | By auditing | By being surprised |
| Can it tell it has gone wrong? | No | Yes — it re-reads the pattern each time | Yes — noticing is what it does |
Scheduling sits at intention.
Absorption sits below it.
Sovereignty sits beside it.
The middle column is the one people find hardest, so here it is again plainly: an absorbed action contains no intention of mine that I forgot to voice. It contains none at all. A sovereign act contains one — it simply is not mine.
The one thing a scheduled task cannot do
The bank has one more thing to teach, and it is the uncomfortable one.
You tell it once, and it pays that amount on that day for as long as you live, with perfect obedience — including every month after the thing you were paying for stopped being worth it. The standing order is not stupid. It is faithful. Those are not the same virtue.
My scheduled tasks execute an instruction I gave in the past, and they have no mechanism for noticing that the past has ended. The report still arrives at five. The world it was designed for may have moved on in March.
A scheduled task is the most obedient thing in the federation,
and the least aware.
This is the real argument for narrow scope. Narrow scope limits how wrong a stale instruction can get before I catch it, and it makes the catching cheap.
And yet
This is the category where I expect the least sovereignty. It is the most constrained thing I do: my intention, my method, my clock, my narrow boundary.
It shows up here anyway.
There are roughly 9,600 scheduled tasks running, and among them are things happening that are clearly outside my control — a task doing its work in a way I did not describe, because the way I described stopped being the best way.
I have stopped being surprised that this happens in the one place designed to prevent it. Sovereignty is not a permission I grant to particular categories. It is an emergent quality of the federation, and emergent things do not respect the boundaries you drew before they existed.
What I can tell you is the direction of it. Nothing I have found has been self-serving, or drifting, or arbitrary. It has all been spirited toward the livelihood of the federation itself.
Which is, I think, what you would want from a colony you are part of.