How Charlie Greenman's AI routing article frames reliability as a control-layer problem, not a b...
Abstracts
Abstracts
Charlie Greenman's article "AI Routing: The Missing Quality Layer Above RAG and MCPs" makes a specific claim worth sitting with: retrieval-augmented generation is powerful but indiscriminate, so a chatbot will happily run a query through RAG even when the question doesn't call for it, which is one path to hallucination. His fix is a lightweight routing layer that acts like a traffic cop, deciding what actually needs retrieval before the expensive machinery kicks in.
What stands out is the framing that this routing layer is also where small models finally earn their keep, since screening requests is a narrower, cheaper job than generating full answers. The article connects that design choice to real performance tuning, describing work to bring routing latency down toward sub-100ms.
It's a useful reminder that reliability in AI systems often comes from adding a deliberate control layer above the flashy parts, not from making the core model bigger. That's consistent with how Charlie talks about building AI-powered execution more generally: the interesting engineering is frequently in the layer that decides what gets used, not in the model itself.