I built a Claude agent that runs Instagram DM ordering for a 7-location sushi chain

I built an AI agent that took over order-taking for a sushi chain with 7 locations. About 90% of their orders come through Instagram DMs, and until now one person typed every reply by hand. How it works: code watches incoming messages through the Meta API and hands each one to Claude (Sonnet 4.6) over the API. The model has a knowledge base with the full menu, ingredients, calories, allergens, delivery zones, hours, prep times and promos for all 7 spots. It talks to the customer for real, helps them pick, explains what is in a roll, flags allergens, and upsells when it fits ("that set goes well with X sauce, want it?"). Once an order is confirmed it pushes straight to the kitchen and writes a record into the restaurant CRM and an admin panel where the owner watches how the agent is doing. Stack: SvelteKit for the site and admin panel, Meta API for the DMs, Claude Sonnet 4.6 for the conversations, pg-boss on Postgres for the job queue, and a CRM integration for the orders. One detail I am happy with: that whole menu-and-rules block has to go to the model on every message, which would normally be expensive. With prompt caching, about 97% of messages read that block from cache at a tenth of the input price, so running Sonnet on every DM ends up cheap enough that the owner never thinks about it. What it doesn't do, by choice: calls, voice notes and photos go to a human. A model guessing at a photo of a handwritten order is how you ship something embarrassing. Plain text handoffs almost never happen, basically just "let me talk to a human," and that is rare. The owner's panel keeps every chat plus the agent's reasoning chain per message, so if something breaks I can see exactly how and why. Still watching quality now that it is live. Happy to answer anything about the caching setup, the Meta API webhook flow, or how the kitchen handoff works.