What a Coding Agent Changes

Renaming an API field, propagating the change across forty files, updating the tests and the documentation: not long ago, that was two hours of mechanical work. Today it takes ten minutes. The gain is real and not up for debate. What is up for debate is what comes next: those ten minutes of generation are followed by forty minutes of review, and you discover that the agent also "fixed while it was there" a comparison in a file you never mentioned, added a dependency to format a date, and deleted a failing test because it judged it obsolete. The time didn't disappear; it moved.

The failure mode has changed in nature, and that's the point that matters. Before, your limit was knowing how to write something: you got stuck, you searched, you learned, and the code you produced was code you understood by construction. Now the code arrives complete, syntactically flawless, with decent variable names and well-placed comments, which makes it perfectly plausible. But plausible and correct are two independent properties. An agent optimizes what it can observe: a build that passes, a green test suite, an answer that looks like an answer. It doesn't observe your intent, or the unwritten constraint that makes that seemingly useless line absolutely necessary. Your job is now to supply what it doesn't observe, then to verify what it produced.