Zach Lloyd, who runs the developer-tools company Warp, says software engineers should stop thinking of themselves as people who write code and start thinking of themselves as people who run a production line. He laid out the case in a long post on X on 29 September, and it is worth reading as an argument, not as reporting.
Lloyd’s starting point is a two-year arc. Engineers went from typing code by hand to steering AI agents from their own machines, and in his telling the grumbling about that step has faded. The next step, he writes, is what he calls a software factory: agents in the cloud handle most of the work from spec to shipped change, and people supervise. He calls the move “inevitable” and admits it may be harder on teams than the last one.
The interesting part is what he concedes. Lloyd says some executives dream of “dark” factories that run with no human oversight, and he warns that chasing them demotivates staff and drives people out. He also dislikes the factory metaphor himself, saying it makes engineers sound like assembly-line workers or like managers of eager but clumsy juniors.
His alternative is a split job description. Engineers keep responsibility for the product: deciding what gets built, how the pieces fit, what good design looks like. Lloyd argues that people still hold more context and better taste than agents. The second job is to improve the factory itself, using engineering measures such as DORA metrics (the standard delivery-speed scorecard) and scoring by a second AI model.
Two practices stand out. First, work happens in public. Instead of tinkering locally and surfacing at code review, engineers prompt cloud agents through tools like Slack and Jira, so colleagues can see the process as well as the result. Lloyd admits that visibility, combined with per-person token counts, can leave people feeling like cogs.
Second, Warp tracks “human touches per PR,” meaning how many times a person has to step in before a change merges. Lloyd’s example is a search, sort and filter feature: a person prompts, adds mockups and iterates on a spec, the agents implement and review the code and record videos of the result, and a person may look once more. He says fewer touches over time signals a better factory.
Lloyd offers one figure. Work needing no human input, such as fixing crashes, small bugs and dependency upgrades, started at roughly 20 to 30 percent of Warp’s workload, he says, and is meant to climb. That is Warp’s own number, from Lloyd’s own post, and it comes with no method or independent check.
The conflict of interest is plain. Warp sells the infrastructure for this model under the name Warp Factories, which Lloyd’s post says is in early access, with qualifying companies offered $10,000 of usage. The essay is both a sincere view of how work will change and a pitch for a category in which Warp wants to be the default. Rivals building cloud coding agents have the same incentive to describe the future this way.
Still, the practical questions land on every engineering lead. Performance reviews at Warp now reward improving the factory as well as building the product, according to Lloyd. That is a choice about who gets promoted, and it is one most teams have not yet made deliberately.
Anyone rolling out cloud coding agents this quarter should decide now what they will measure per engineer, and who will see it, before the dashboards decide it for them.
Argued by Zach Lloyd in his own post on X on 29 September 2026.
