I think that makes sense for now in individual contributor cases, but the turnkey solutions have gaps when it comes to the most powerful capabilities and I don't see harnesses going away any time soon for that.

E.g. you can't use coworking/projects for shared memory in most domains because this hinges on compacting/refactoring true reusable facts and organisational knowledge (as opposed to code, schemas, technical references). The heuristics to find and decide on candidates for memory in this case are quite specific to the institution you work in. For example, the boundaries of the domain of "sales" versus "marketing" vary lots from place to place, as does the trustworthiness of certain staff, the way they they communicate or make decisions that contain the knowledge you want to store to memory, and so on.

I'm finding harnesses are still vastly better for grabbing this sort of complex institutional knowledge. Also, they're just a bit of a safety net for any frontier labs going bankrupt in the next couple of years, if we need an easy way to port over to open weights.