Port the site. Keep the conductor.
I moved my whole site between hosts this week. The model that planned the move never executed a single task of it.
Fable planned it, sitting at the gate. The migration itself, the wrangler config and the adapter and the DNS cutover, was executed by a swarm of cheaper models, each routed to the one slice of work it could do.
Rent the conductor. Swap the orchestra.
The conductor is the harness and the routing. The orchestra is the model, and it is swappable. Here Fable held the baton and budget lanes played the parts. Elsewhere the conductor seat itself is a cheap large-context model: one is running 34 parallel workers for me right now, another wired the grok fleet that shipped ten landing pages overnight.
The orchestra (model)
Agents in the fan-out: 235
Real published rates, Jun 2026 ($/1M in/out): Claude (Opus 4.8) $5/$25, GLM-4.6 $0.43/$1.74, MiniMax M2 $0.26/$1.00. Assumes ~56k tokens per agent (45k in, 11k out).
The intelligence is the routing, not any single model.
Two caveats, honestly. It was a day and a morning, not a day: preview green before dinner, production flipped the next morning. And I sat at every gate. Nothing ships itself.
The port is not the point. The point is the expensive thinking and the cheap doing have been peeled apart, and you route each to whichever model earns it.
What is the one job you would hand to the cheapest model that could do it, if you stopped insisting the smartest model do everything?
//A<3
Build it today
The two artefacts below are the load-bearing parts of this build, redacted of secrets.
↓wrangler.jsonc (redacted) ↓dispatch-matrix sliceThe pattern they encode: one conductor (the harness, the routing, the gates) and many swappable workers (the cheapest model that can do each slice). The conductor never moves. The orchestra is rented by the hour.