The split you describe, dedicated developers for deep product integration and an agency for a broad proof of concept, holds up. I would add a third case that comes up often: a single, well-defined problem that needs to be solved properly once, where neither new headcount nor an open-ended engagement fits.

For that case the useful question is less about who builds it and more about how the work is bounded. One outcome stated in a sentence, acceptance checks written before the build, the repository in the client's account from the first day, and a fixed price against that outcome. If a vendor cannot write that page, paying them by the hour does not fix it. If they can, the hire or agency decision gets much easier, because you are finally comparing like with like.

Hiring tends to make the most sense once the problem stops being one problem and becomes a product roadmap.