AGENTS AT WORK · 09 A visual series

Right-sizing the model

Click each task and watch it route. Mechanical work goes to a small fast model; judgment work goes to a strong one.

Small fast model
$0.000
Strong model
$0.000

The problem this replaces

Every task, one expensive model

EVERYTHING RUNS ON THE STRONG MODEL $0.480 — for 3 checks that never needed judgment

A compile check costs the same as writing the whole test suite. Nothing is cheap enough to run constantly.

How it works

One fork decides the lane

A TASK NEEDS JUDGMENT? yes no STRONG MODEL FAST MODEL
Mechanical, repeatable checks never need the expensive lane — only the ones that require weighing tradeoffs do.

This is also why impact analysis and performance review can run on every single change — they're cheap enough to never skip.

See it in action

One run, two bills

fast: impact analysis, lint, compile → $0.006 · strong: plan, review, tests → $0.24

The mechanical checks ran on every commit. The judgment calls ran once each.

So what

The boring checks are cheap, so they can run every single time.