Tim Davis, CEO of Modular, has been running an experiment. He built Compound Loop, a system that chains multiple AI models together to write, review, and merge code while he sleeps. Davis would set it running on a real problem before bed, then wake up to triage a stack of pull requests that didn't exist the night before. "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," he writes. The system doesn't just complete tasks. It compounds. By 8 a.m., Davis wasn't catching up on yesterday's work. He was deciding which overnight jobs to keep while the system kept generating more.
The implications for engineering roles are messy. The optimistic story says everyone levels up. That's not what Davis sees. Some engineers do move up, becoming system architects and product managers working at higher abstraction layers. But others get pushed down into spec writing or reviewing machine output or what Davis calls "agent babysitting." He's blunt about the stakes: "These fragmented roles will be paid less, valued less, and in many cases become career dead ends." The pay gap between top-tier operators running agent fleets and middle-tier workers managing their exhaust, he argues, will be wider than the old gap between engineers and sales reps.
Davis applies Jevons paradox to code. In 1865, William Stanley Jevons noticed that more efficient steam engines led to more coal consumption, not less. The same pattern holds now. As the cost of writing code approaches zero, teams don't write less code. They write vastly more. The constraint shifts from how fast engineers can type to how fast product surfaces can absorb the output. Davis says the best AI-native teams are shipping three to ten times what they shipped a year ago, and the curve is still bending upward.
But more code doesn't mean better code. When supply explodes, knowing what to point the agents at becomes everything. Filtering what comes back and integrating it into something coherent, that's the scarce skill now. Production is cheap. Selection is hard.
And the old moats haven't entirely disappeared. Kernel performance, compiler design, and hardware abstraction remain defensible because those domains still require determinism. The probabilistic future hasn't reached all the way down the stack.