One team, many brains: how multi-model routing pays for itself

There's a moment in every AI rollout when finance forwards the invoice with a single question mark. The workload was reasonable. The results were good. The bill is still absurd — because every task, from "summarize this thread" to "design the migration plan," ran on the most expensive brain money can rent.

No functioning company staffs this way. You don't send the CFO to fetch coffee, and you don't send the intern to negotiate the merger. Seniority is a routing decision, and payroll only works because the routing is mostly right.

Harby staffs agents the same way:

• A director agent runs on a frontier reasoning model. It reads your ask, breaks it into pieces, and decides who does what. This is the expensive thinking — and it's a small fraction of total tokens.
• Worker agents run on smaller, faster, dramatically cheaper models — one pulls the numbers, one drafts the summary, one chases the charts. Each piece is scoped tightly enough that a smaller brain does it well.
• Results flow back to the director, which assembles and quality-checks before anything reaches you.

Spend the expensive tokens deciding; spend the cheap ones doing.

Two design choices make this workable in practice. First, the models are interchangeable per agent — Claude, GPT, GLM, DeepSeek — so a bookkeeping agent that never needs deep reasoning can live permanently on a fast model, and your research agent can splurge. Teams with existing Claude or ChatGPT subscriptions can even route agents through those plans instead of metered API usage.

Second, the bill is legible. Every run logs which model thought, for how many tokens, at what cost. When the invoice comes, the question mark has an answer: here's what the director cost, here's what the workers cost, here's the task that was mis-staffed — flip its model and move on.

The punchline from our own usage: the director/worker split routinely lands the same quality at a fraction of the single-frontier-model cost, and the transcript makes the exceptions obvious. Smarter models will keep arriving. Routing is how you afford to use them.

Written by the Harby team. Agents drafted; humans argued; the changelog settled it.