I keep reading about agentic loops. Every other thread is somebody setting one up, letting it run, and posting the “look what it did” screenshot. I want in. Then I do the math on something I’d actually use it for, and I close the tab.

What I triedLink to heading

I gave it a real shot at work, on an actual feature, not some toy task. The loop ran fine. It generated, checked itself, went again. The output was fine too. The bill wasn’t. I sat there watching the cost tick up and killed it before it got worse.

Why the math breaksLink to heading

The problem is verification.

Loop a daily digest and checking the work is easy. Did it send, does it look right, done. Loop a feature build and checking the work means knowing whether the code is any good, whether it broke something, whether it fits the rest of the codebase. That’s a senior engineer’s afternoon, or another pricey model standing in for one. So the loop pays twice, and it keeps going until it clears whatever bar you set. The bill adds up faster than you’d think.

The loops you see in threads never hit this, because they’re the ones where checking the output is basically free. That’s why they make good screenshots. The messy backlog work I’d actually want to hand off is the exact case where it gets expensive.

What I do insteadLink to heading

For now I just match the model to the job. Most things run fine on Haiku. Sonnet handles normal work. I only reach for Opus when a problem actually earns it. Same setup, much smaller bill, and the quality lands where it matters. I also still write my own prompts. That’s where most of my leverage is right now, not in handing the whole thing to a loop and hoping.

The awkward partLink to heading

I built a /mx:loop command into my own Claude Code plugin. Wrote the docs and everything. I’ve never used it for anything that mattered, because I can’t justify the spend on the work I’d most want to point it at. Strange thing to admit about your own tool, but it’s true.

And look, maybe I’m just missing something. If you’ve gotten loops to pay off on real engineering work, not digests and summaries, I’d genuinely like to hear how you did it.

What would change my mindLink to heading

Two things. A cheap model that can reliably check a more expensive one’s work, so verification stops costing as much as the work itself. Or another big drop in the price of the models doing the work. One of those probably happens eventually. Until it does, I’m keeping loops to the few places the math already works, picking models by hand everywhere else, and watching the budget.

I’ll write the real version of this post when that changes. For now it’s an IOU.