Non-coder here — building a pipeline whose whole point is using frontier intelligence as cost-effectively as possible. Does this make sense
I'm not a developer — I built (two) things (so far) by directing AI agents. The goal of my pipeline is simple to state: pay frontier rates only where judgment matters, cheap rates everywhere else.
How it's split: a frontier model designs the blueprint, DeepSeek/GLM Flash writes the code, a different frontier model does ONE review pass, and a cheap model repairs whatever it flags. One review, no further loops. Nothing ships without my approval.
I got burned once paying for one model and getting silently served another, so every seat's
route is verified against what was actually asked for.
The pieces exist separately — MetaGPT does roles, RouteLLM does cost-splitting — but I haven't seen the combination. So, honestly: does this make sense, or am I overcomplicating it?