Hot take: AI coding agents aren't making senior developers faster

I've started wondering whether AI coding agents are actually improving developer productivity at the senior level, or whether they're just moving the work to a different part of the process.

For smaller tasks, the productivity gain feels obvious. Generate some boilerplate, write tests, refactor something repetitive, investigate an unfamiliar API — agents are great at that.

But once the task involves an existing codebase with a lot of context, things get more interesting.

The agent has to understand the architecture, figure out which files actually matter, make changes without breaking unrelated behavior, and then explain why it made those changes.

At that point, I sometimes spend almost as much time reviewing, correcting, and steering the agent as I would have spent implementing the change myself.

And there's another problem: the better the agent gets at producing code that looks reasonable, the harder it can be to notice subtle architectural mistakes.

So I'm starting to think the real bottleneck isn't code generation anymore. It's context + verification + supervision.

Maybe the productivity curve looks something like this:

Junior developer + agent → huge boost Senior developer + agent → depends heavily on the task Complex production system + agent → supervision becomes the bottleneck

I'm curious what others are seeing in real projects.

Have AI coding agents genuinely made you faster overall, including review/debugging/cleanup, or are they mostly making the "first draft" of the code faster?