Work / Lab
Model routing in practice
Routed models by task, cost and vendor across three apps; the routing worth keeping is the rule for who checks the work.
- TypeScript
- Rust
TL;DR
- I’ve routed models by task, cost and vendor in three codebases, each built by AI coding agents under my direction.
- A research workflow I had map Concord’s market rated routing on its own as table stakes.
- The rule I keep is the verification one: the model that checks the work comes from a different provider than the model that did it.
Three apps, three kinds of routing
- JobApp routes by task tier: judgment goes to Claude Opus, extraction to Sonnet, classification to Haiku, with per-model pricing and a cost estimate for every call.
- Redl routes by where a model can run: each council seat goes to a local or a cloud model, and it recommends offloading when a local model won’t fit in memory.
- Concord routes by capability, price and vendor: it scores models on both, and seats each challenger on a different provider from the analyst it checks.
What I concluded
Then I had a research workflow map Concord’s market. It rated routing by itself as table stakes, and the two claims that survived were a blocking citation tie-out and cross-vendor challenge with dissent preserved. So cost routing is plumbing. The routing decision that changes outcomes is who gets to check whom.