A Salesforce research engineer runs five or six AI coding agents at once: Reveneau

Why it mattersThe post shows how one engineer spends about 3,000 dollars a month and still rewrites 4,000 lines by hand. Review becomes the bottleneck, not writing.

Ben Stolovitz, a research engineer at Salesforce AI Research, wrote up a year of using AI at work. The post reached 87 points on Hacker News within hours of going up. Stolovitz last published a similar write-up a year ago. He says "everything's completely different" since then.

The piece is useful because it carries real numbers from one engineer's practice. It names tools, it names prices, and it names the parts that still do not work.

What the practice looks like now

Stolovitz runs five to six Claude Code agents at the same time. Each one works on a feature, a bug, or a pull request in its own worktree. He pays for Claude Pro at 17 dollars a month and GitHub Copilot Pro at 10 dollars a month. The rest of his spend goes on API calls. The total, he writes, is about 3,000 dollars a month.

A prototype that would have taken him a week before, he says, now takes about 20 minutes. He gives one example where Claude Code added noise reduction to his open-source audio project in a single prompt. He calls out autocomplete as "still good", with a screenshot of it offering several dim lines.

What still does not work

Stolovitz rewrote 4,000 lines of AI-generated code by hand. He says the rewrite was more flexible and simpler than what the agents produced. The post frames this as the central limit of the current wave. The agents can write code that compiles and passes tests. A person still has to rework it, because taste and experience do not sit inside the model.

AI writing, he says, is "plodding, repetitive, and vapid". He will not use it for personal messages. He also flags a concern about research: a language model, asked to find papers supporting a claim, can usually find some. That makes it a weak tool for a literature search. The claim comes out verified when it should have been tested.

What a team can take from it

Three things, each tied to a number in the post. The first is that running several agents at once is a real pattern, not an experiment. Five to six parallel agents is what one engineer is using this week.

The second is that cost scales with practice. 3,000 dollars a month is one engineer. A team of ten at the same rate is 30,000.

The third is that the rewrite rate is the number worth measuring. Review is where the work moved. A team that tracks rewritten lines sees the shape of its AI work fast. A team that tracks lines produced sees less.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.