In my own experience this is 180 degrees from reality. As a generalist, feeling out the depths of a single domain (something I've been forced to do at least 50 times in my career, to the point that I'm probably a global expert in at least 2-3 things I don't actually care about, but are poorly documented and not especially lucrative on their own) is something that's basically a bunch of Google searches, reading source code, and writing/running tests manually, none of which I really care about short of getting to "the right solution."

Meanwhile, as a generalist who has a basic understanding of general things, everything from how to design efficient network protocols, to how cache lines affect the performance of sorting algorithms, without being a real expert in any of those things, I act as a constant course correction for AI agents doing work on my behalf, in a way that LLM context windows simply cannot replicate.

To give a concrete example, I recently used agents to build a specialized sync protocol that broadly resembles Dropbox. It's nowhere near as efficient in terms of how blocks are synced (because it entirely happens on a LAN and the cost difference is minimal), but I constantly had to make objectively more valuable course corrections on how the sync actually traversed the participating nodes. If I'd just let the LLM drive, it would have come up with a reasonably efficient algorithm (better than I probably would have done on my first try in the same timeframe) that would have had an obvious (to me) single bottleneck.