Tim Davis, CEO of Modular, built a side project called Compound Loop. It uses frontier AI models to write and ship code autonomously, handling the review process along the way. He'd set it running before bed and wake up to a stack of pull requests. Some were excellent. Some were flat wrong. And some surfaced questions he didn't even know to ask. "For the first time in the history of knowledge work, the person who went home did not take the only copy of their brain with them," Davis writes. He calls this the 24-7 employee. While you sleep, AI agents work in parallel, and the compounding is relentless.
The reality for engineering teams is messier than the "everyone levels up" narrative. Davis sees a split. Top engineers become more effective product managers and architects, working at higher abstraction layers. Others get pushed into what he calls "spec writers and agent babysitters," translating intent into prompts and grading machine output against standards they might not fully possess. Davis predicts the pay gap between top-tier engineers running agent fleets and middle-tier workers managing their exhaust will exceed the old engineer-sales gap. He's already watching it open inside companies he tracks.
Davis applies an 1865 observation by economist William Stanley Jevons to software. Jevons noticed that more efficient steam engines led to more coal consumption, because efficiency expanded what was worth building. Same logic for code. As the unit cost of writing code approaches zero, teams write vastly more, and the best ones ship three to ten times what they shipped a year ago. The bottleneck shifted from typing speed to how fast a product can absorb output. But Jevons had a second lesson: when supply explodes, selection becomes everything. Cheap code without judgment is waste.
The most valuable work in software right now is pointing agent fleets at the right problems and filtering their output into something coherent. Production is solved. Direction and selection are where it gets hard. Davis is honest about the tradeoffs. You no longer know your code works. You believe it works, with a probability you can't precisely state. Deterministic engineering still matters at the lowest levels like kernels and compilers. But the higher layers are becoming probabilistic, fast.