Eight months of daily use. Three or four pull requests shipped in a single day. Tests passing. Code merged. And still, a nagging question: did I actually work today, or did I just supervise?
This is what senior developers are sitting with as AI coding tools become a daily driver rather than an occasional shortcut. It's not the job-security question - that one has been argued to exhaustion. It's something harder to articulate: a shift in what "doing the work" feels like from the inside when the output is indistinguishable from before.
What Actually Changed in the Workflow
The traditional developer day had a rhythm: understand a problem, break it down, write code, debug, refactor, review. The cognitive load was distributed across all phases. The satisfaction - the thing that makes development feel like a craft - came partly from solving the hard parts directly.
With Claudee Code](/tools/claude-code/) doing most of the implementation, that rhythm shifts. You spend more time on the front end (defining what you want clearly enough that the AI can execute it well) and on the back end (reviewing output, catching errors, making judgment calls, rejecting subtly wrong architectural decisions). The implementation phase, where a lot of the craft traditionally lived, is mostly delegated.
This isn't supervision in the way a manager supervises. Catching the bugs that look right, recognizing when the agent's architectural choices are subtly off, writing prompts that specify the right constraints - all of this requires genuine technical depth. But it is a different kind of work than writing the implementation yourself.
Why Senior Developers Feel This More
A junior developer using these tools might feel productive and not question it much. A developer with a decade of backend experience has a calibrated internal sense of what "hard problem" feels like from the inside. When that sensation is absent from a day that was, by every external measure, productive, it registers as something off.
The honest answer is probably that the feeling of craftsmanship was always somewhat distinct from the output. You can write elegant code that no one uses and ship messy code that solves a real problem. AI coding tools are forcing a separation between those two things that good developers had managed to keep together.
What remains the human's job - and where years of experience still matter - is judgment: knowing what's worth building, catching when the AI is confidently wrong, and maintaining the conceptual model of the whole system while the agent handles the individual pieces. That's real work. Whether it satisfies in the same way as writing the implementation yourself is a different question, and probably a personal one.