Useful writeup - and new term for me. What's interesting is the RFT autofixer, and finding better ways to stay on frontier models while injecting appropriately new, specific context (like documentation). The pattern I used in the 3.5-turbo days was separate tool calls with individual tools being separate model calls (we didn't have the term subagent yet) where additional documentation was injected. So the flights tool would be 3.5 + GDS docs, the visas one would be 3.5, etc. Another pattern I've heard of in db land is using embeddings to semantically find the closest complex queries and use that to enrich knowledge. The tradeoff here is simple: If you use SFT/RL, you'll be behind on the latest models, but if you get it right you free up context and get better intelligence. If you stuff context, you can do a lot but you lose context length and risk few-shotting models down the wrong paths.
Under the hood of v0’s composite model. v0 is built with custom models and knowledge tuned for React and web development, paired with pluggable, state-of-the-art frontier models. Here's how we build and evaluate it. vercel.com/blog/v0-compos…