#7 Switching from French to Italian

Converting a working AI agent from one purpose to another — say, customer support to code review — isn't swapping a file. Five out of seven components have to be redesigned. Only the base infrastructure stays. Everything built on top of it gets replaced.

The instinct is "just change the prompt and connect different tools." That's like saying "just swap the recipes" when you're converting a French restaurant to Italian. Recipes are one of five things that change. And they're not even the hardest part.

The French Restaurant That Works

You run a French restaurant. The system is dialed in. A French wine importer picks great bottles every week. A truffle farm delivers seasonal ingredients. You've worked with the same Normandy butter supplier for ten years. The recipe book is packed with years of proven dishes — bouillabaisse, coq au vin, crème brûlée, soufflé. Prep cooks have clear roles: French sauces, pastry dough, consommé. The brigade is tight.

Then the owner says: "What if we switch to Italian?"

What Actually Has to Change

New recipes mean new ingredients. Instead of the truffle farm, you need a San Marzano tomato farm. Instead of the French wine importer, an Italian one. The entire supply chain gets rebuilt. Prep cook roles change — no more French sauce specialist, now you need someone making fresh pasta and someone prepping antipasto. The brigade gets restructured for Italian-style service. And the chef's identity shifts from "I'm a French chef" to "I'm an Italian chef."

Only two things stay: kitchen equipment (ovens, refrigerators, stoves) and the chef's fundamentals (knife skills, heat control, palate for balance). These work whether you're cooking French or Italian. Those two constants are actually why the switch is possible. If you had to replace the ovens too, you'd be better off opening a new restaurant.

Five out of seven change. Only two stay.

Same Pattern, Different System

Converting a customer support agent into a code review agent follows the same structure.

Base infrastructure (LLM capability, MCP protocol) stays. Everything on top (connections, skills, subagents, teams, prompt) gets redesigned.

"Can't One Agent Do Everything?"

That's like thinking a French chef can probably do Italian just fine. A chef with strong fundamentals adapts faster, sure. Ask a French chef to cook Italian and they'll manage. But to run a proper Italian restaurant, you need Italian recipes, Italian suppliers, and a team structured for Italian-style service.

Same with AI. A smart LLM given "review this code" will produce something. But a proper code review agent requires GitHub integration (MCP), security and convention checklists (Skills), lint and test helpers (Subagents), and a frontend/backend/security reviewer team (Agent Teams) — all designed from scratch.

An agent is not a general-purpose tool. It's a purpose-built system. When the purpose changes, you redesign the system. The ovens stay. Everything else gets rebuilt.