Back to blog

1 juni 2026

Claude Cowork rethinks the task every time. That’s where the cost is.

An AI agent like Claude Cowork does your task and it feels like magic. But every time it runs, it reasons the whole thing out again, and that costs tokens each time. For a recurring task you pay that thinking bill over and over. Building it once is far cheaper in the long run.

Gaëtan Steijaert
Written by Gaëtan Steijaert Software architect from Ghent

You hand a task to an AI agent like Claude Cowork, and it sorts it out. It reads along, thinks, decides and acts. It feels like magic. So you let it do that task again. And again. Every day.

What I saw

Every run, that agent starts from scratch. It reads everything again, weighs it, decides and acts. That thinking costs tokens, every single time. For a one-off, that’s nothing next to the time you’d spend yourself. But for a task that comes back every day, you pay that thinking bill over and over.

And that’s the sting: the cost grows with how often you run the task, not with how much it earns. The more the agent impresses you, the more you use it, the faster it adds up. For the boring, recurring work you pay the price of original thought every time, while the answer is the same every time.

Claude Cowork do it in the chat each time thinkthen act thinkthen act thinkthen act Every run the agent reasons it all out again. Tokens every time, cost every time. Building with GA.ST think it through, once think it through(with AI) build runs, runs, runs Think and build once. After that it just runs, near-free per run.
Reason it out every time, or think it through once and then let it run.

What I do

For a task that comes back and that we understand, I build it once. The thinking I do with AI, because that’s where it’s strong: finding the pattern, catching the edge cases, writing a first version. But that result I then cast into code. A fixed chain that simply runs. No fresh thinking per run, near-zero cost per run, and it does exactly the same thing every time.

For invoices, declarations and syncs, that “exactly the same every time” is a feature, not a limit. You don’t want an agent there that rethinks it each run and occasionally makes a different call. You want predictable, and you want cheap.

I’m not throwing the AI out, quite the opposite. I use it where it shines: the new, the messy, the judgment calls, the one-offs. For the boring repeat, code wins, on price and on reliability.

A rule of thumb

Does a task come back, and do you understand it well? Build it once. Is it new, messy or rare? Let the AI think it through. The art is in knowing which of the two you’re looking at. And that’s exactly where I’m glad to help you look.

And for you?

Asking an AI the same thing every day? Then maybe that task just needs to be built once, so it runs on its own after that. Let me come and look. The first look is free, no strings.