How To (Not) Spend $10k/wk on Coding Agents - Allen Pike

Allen Pike

Articles

About

Follow

How To (Not) Spend $10k/wk on Coding Agents

On too much of a good thing.

June 30, 2026 •<br>8 min read

Word on the street is that the cost of building software is going to zero.

Zero? Sounds like a good deal!

Over the last year, my co-founder and I have iteratively automated our coding loops. Each time better tools revealed a bottleneck, we’d address it. Agents would sometimes break things or propose slop, so we added more tests and guardrails. Our PRs piled up, so we automated the mechanical parts of review. When UX review became a bottleneck, we had agents attach demo videos to PRs.

Nothing revolutionary – just stubborn hill-climbing on our dev loops.

By April, Jenn had our agents humming. They were automatically and safely fixing lints, nitpicks, merge conflicts, outdated dependencies, and other maintenance chores. Our velocity kept increasing, and I started to file bugs and propose improvements that we previously wouldn’t have had time for. We could respond to user requests same-day. We were working hard, but moving fast.

It was also generating a lot of Anthropic and Cursor and OpenAI overage emails?

But this is the way of the future. We’re a funded startup. And coding costs are supposed to go to zero. And Jenn’s shipping 30 PRs in a day now. Being delightfully responsive to our customers is a superpower!

Okay okay, this is a lot of overage emails. I’ll assess our weekly spend.

Hey Jenn? How surprised would you be if I told you we’re spending $10,000 a week on coding?

I have been working a lot. But… that is too much.

StrongDM infamously claimed in February that you should be spending at least $1,000 per engineer per day on your coding factory. Well, we did it. Do we get a prize for our profligacy?

This is too much

Yes, our coding agents were delivering a lot of improvements. But you can hire a pretty good engineer for $10k/week. Within a couple days we’d cut this spend way down, while maintaining most of the velocity – some techniques for this below.

But every week, I hear of more teams hitting this same transition point. Coordinating agents to do something is less compelling if they’re more expensive than having a human do that same work. Thus, we’re moving from the era of “how can we use more coding agents?” to “how can we get the most out of our coding agent spend?”

This change is all around us. Sam Altman has started saying that token costs are a “huge issue”. Brian Armstrong is talking about how to do more agentic coding at a lower cost. Uber is throwing model labs under the bus over agent spend.

The vibe has shifted. There are a few reasons for this, but much of it is downstream from cloud coding.

Cloud coding costs

While you can go pretty far juggling a few agents on your laptop and leaving it propped open all day, getting the most out of coding agents really requires them to run in the cloud. We’ve seen a lot of workflow benefits from cloud coding, many of which are described in OpenAI’s Symphony coding factory blog post from April.

However, it’s worth noting this in OpenAI’s writeup:

This way of working dramatically reduces the cognitive cost of kicking off ambiguous work. If the agent gets something wrong, that’s still useful information, and the cost to us is near zero. We can very cheaply file tickets for the agent to go prototype and explore, and throw away any explorations we don’t like.

“Very cheaply” if your tokens are free! Model labs’ employees are shielded from the cost of full-on software factories.

For the rest of us, cloud coding gets expensive for the same reasons that cloud compute gets expensive:

It makes it easy to do lots of work at once

It’s more expensive per unit of work than using your own laptop

It can make wasteful work go unnoticed

When we first got into cloud agents, we ended up primarily using Cursor’s. While Codex and Claude Code are the current monarchs of local development, Cursor’s cloud coding harness and workflows are quite a bit more mature than Codex’s or Claude’s.1 However, this capability comes at a premium: Cursor charges a markup on API costs on top of Anthropic and OpenAI, and you can’t turn off their “MAX” mode for cloud coding. Using Claude Fable for Cursor Cloud Agents is fecking expensive.

At the AI Engineering World’s Fair this morning, Openclaw creator Peter Steinberger gave a short talk about cloud coding workflows, where he joked about token costs:

Last year, I was primarily constrained by tokens. And I fixed that! By joining OpenAI.

Luckily, there are other approaches to deal with exploding inference bills. Coding costs are the multiplicand of token cost and token count, so let’s look at each.

Cheaper tokens

While the typical advice for coding has long been to use frontier models for everything, and model labs have been mostly focused on those most-expensive models (with Claude Fable hitting $10/Mtok), that’s starting to...