If code gets cheap, defining the right product becomes the scarce work

Laurie Voss’s essay, shared by Simon Willison, starts with two explicit assumptions: agents eventually handle much of the software lifecycle, and demand for useful software has no practical near-term ceiling.

On those assumptions, the scarce work becomes discovering what a customer actually needs, defining success precisely and making the result pleasant to use. Requirements are specific to particular organizations; making implementation cheaper does not automatically reveal them.

The sharpest consequence concerns training. If junior engineers no longer learn through small implementation tasks and senior review, companies lose an established route for developing judgment. Hiring experienced people cannot permanently substitute for producing them.

The argument shifts attention from generating more code to building the capacity to decide what deserves to exist and whether it works for its users.

The caveat is substantial: Voss acknowledges thin evidence that agents can reliably take over operations and scaling. His forecast also assumes product discovery remains comparatively resistant to automation. It is a useful conditional model of the profession’s future, not an established result that programming jobs have ended.