Agentic coding collapsed the cost of writing software. It did not collapse the cost of specifying it, verifying it, or judging whether it should exist. That inversion — cheap building, expensive judgment — is the whole story of how engineering is changing.
The scarce input is judgment
When an agent can execute almost any well-specified task at near-zero cost, the value moves entirely upstream: to deciding what the task is, what done even means, and whether it should be attempted at all. The work that's left is the work that was always hardest — and it doesn't get cheaper.
So the senior engineer's job inverts. Less time writing code, more time defining what correct means: the constraints, the invariants, the boundaries. Judgment, encoded — and applied at machine speed.
What changes in practice
Estimates change shape: implementation time stops dominating, so the question becomes how many decisions a feature contains, not how many lines. Review changes shape: you audit the spec and the verification, then spot-check the code, because reading every generated line is neither possible nor useful. And ambition changes shape: features you'd never have budgeted — a live entity graph powering a marketing site, say — become afternoon projects, which moves the bottleneck to taste.
The discipline that survives is the oldest one: know precisely what you want before you ask for it. Agents just raised the price of not knowing.