The decide-execute-deliver sandwich, or why coding agents haven't replaced engineers

Narayanan and Kapoor’s latest extends their “AI as normal technology” thesis to the case everyone keeps invoking — software engineering — and it’s the measured counterweight to the existential-threat framing. Their frame is a “decide-execute-deliver sandwich”: AI genuinely compresses the execution layer, the actual writing of code, while the layers on either side — deciding and specifying what to build, then verifying and being accountable for what shipped — resist automation for structural reasons rather than mere capability gaps.

The numbers they cite make the point sharper than the argument alone: an 8x increase in lines of code produced, but only about 30% more releases. Execution got cheap; the bottleneck moved to everything execution was never the hard part of. And the distinction they draw between “vibe coding” — unsupervised, low-accountability — and “agentic engineering” under human supervision is the one that actually predicts who keeps their job.

I find this genuinely clarifying for thinking about my own work, where the spec and the accountability — does this signaling behave correctly, can I stand behind it — were always the real labor and the code was the easy mile. It’s a calmer, better-grounded read than most of what’s circulating, and the kind of argument worth keeping around to point people at.