Why we route tasks, not models

Every "AI gateway" pitch sounds the same: send us your prompt, we'll pick the cheapest model that can answer it. It's a reasonable thing to build. It is also solving the easy 10% of the problem.

Here's the 90%: your best engineer reviews a pull request completely differently than a new hire would. Your best support rep writes a refund reply that quotes the policy, your newest one doesn't know the policy exists. The gap between your best and your average AI output has nothing to do with which model answered. It's the thinking that happened before the model was ever called.

A task router asks a different question

A model router asks: "given this prompt, which model is cheapest and good enough?" A task router asks: "what is this person actually trying to do, and what's our playbook for it?" That second question is the one your AI team actually cares about, because the answer determines the model, the effort level, the instructions injected before the model sees anything, and whether the task gets broken into subagent steps.

In Verybear, that answer is called a flow. "Review pull request" is a flow. "Draft customer reply" is a flow. Each one pairs a plain-English description with a model, an effort level, a playbook (the way-of-working instructions your best person would give a new hire), and, where it helps, a breakdown into subagents. The person typing "review PR 1182" into Claude or Cursor never sees any of that. They just get the output your best reviewer would have produced.

Why this matters more as usage grows

Model routing optimizes a number that gets less interesting as your AI usage matures: cost per call. Task routing optimizes the number that actually keeps a CFO or a Chief AI Officer up at night: consistency. When 200 people are typing requests into five different tools, the question isn't "did we use the cheap model." It's "did everyone get the same good answer the first time."

There's a second benefit that compounds: when a flow improves, everyone using it improves at the same moment. Nobody needs to be retrained, nobody needs a new prompt library emailed to them. The flow syncs to every desk (Claude, Codex, Pi, Paseo, Cursor, VS Code, wherever people actually work) the next time they open it.

We still pick the model. We still pick the effort level. But those are outputs of understanding the task, not the whole job. The whole job is making sure the way your company does a thing in AI is the way your company actually wants it done, every time, on every desk, without anyone having to remember a prompt.