Skip to content

The Car Metaphor

coo1white edited this page Jun 8, 2026 · 1 revision

The Car Metaphor

A plain-language way to lock down what Cool Workflow (CW) is — and, just as importantly, what it deliberately is not.

CW is a small car with no engine of its own that still needs fuel. In today's AI world the engine-maker and the fuel-seller are the same company, so CW lets you swap both the engine and the fuel grade to fit the job.

Three roles — don't conflate them

Role Maps to
🚗 The car CW: chassis + dashboard + flight recorder + transmission + (from v0.1.38) autopilot
⚙️ The engine The model itself (Claude / GPT weights + inference) — the part that turns fuel into motion
⛽ The fuel Tokens — the metered consumable you pay for and burn

The engine is not the fuel. They are two different things.

What's strange about the AI world

In the real world Toyota builds engines and an oil company sells fuel — two separate businesses, and any engine runs on any station's fuel.

In AI it's inverted: the same company builds the engine and sells the fuel, and each engine only runs on its own fuel.

  • Anthropic builds the Claude engine (closed) and meters fuel by the token.
  • You cannot run the Claude engine on OpenAI's fuel — engine and fuel are bundled.

So each AI provider is an integrated engine-factory + proprietary filling-station.

The red line: build the car, never the engine or the fuel

Building a frontier engine costs billions in training and is closed-source — you can't out-build them, and you shouldn't try. CW's choice: touch neither the engine nor the fuel; build only the car that any engine and any fuel can drop into. Claude today, GPT tomorrow, local Ollama the day after — the car is yours, the engine and fuel are swappable (this is what backend-neutrality means; see Architecture Principles and Runtime Contract).

The moment CW refines its own fuel or builds its own engine, it stops being the neutral car that runs on anyone's fuel and becomes just another fuel-seller — throwing away neutrality, its single biggest advantage.

v0.1.38 = adding autopilot to the car

  • Before: a manual-shift car — you steer every worker by hand (a human writes each result.md).
  • v0.1.38: cw run --drive — set the destination and the car drives the whole route itself: plan → dispatch → fulfill → verify → commit. The fuel (model) is still poured in from outside, the flight recorder still records evidence the whole way, and on failure it pulls over (fail-closed park) rather than fabricating arrival.

Why an open-source car is itself the selling point

A flight recorder is only trustworthy if you can open the hood and read its wiring yourself — only an open-source car makes "trust the black-box recorder" coherent. A closed-source audit layer is a contradiction. And it's a small car (small kernel, glass hood), not a LangChain-style behemoth with the hood welded shut.


Related: Architecture Principles · Runtime Contract · Distribution Strategy

Clone this wiki locally