Does a strong ICM layer make new model releases less important?

If your system already does context management, retrieval, memory, and orchestration well — does upgrading from Model A to Model B actually move the needle? Or are we hitting a point where the system around the model matters more than the model itself?

Personally, I’ve started caring less about which model is “best.” I often run two models at once when I get stuck on something — not because one is smarter, but because I end up acting as a mediator between their ideas, pulling the best parts from each to solve the problem. The model feels less like the bottleneck and more like one input into a process I’m steering.

Curious where you land

Are you still upgrading models the day they drop, or has that urgency faded?

Has better context/retrieval ever made a “worse” model outperform a “better” one in your stack?

Anyone else running multiple models side by side and playing mediator like this?

10 comments

Does a strong ICM layer make new model releases less important?