Agentic operations
Every app, service and agent in our system holds a scoped, signed, revocable capability, and the machine that acts verifies it before acting. That is the whole authorization model, and it is why an agent can work for a factory without owning its data.
Agents draft quotes, questions to suppliers and follow-ups. Every outbound message to a real business goes through an approval-gated outbox where a person reads it and presses send. An agent that can send one unsupervised is an agent that can order 400 parts instead of 4, so none can.
Mail, Messages and WeChat arrive in one inbox that agents read through the same service, so a supplier's answer on any channel lands next to the question that caused it.
Orders, shipments, operations and scenario runs are readable over HTTP with the same observed, simulated or unknown label the screen shows. An agent asking the twin where an order waits gets the same answer a person does, and the same honesty about its source.
The factory agent network lets a customer's software and your office exchange RFQs, quotes, orders, change notices, shipments and gate events as signed messages. Both sides keep the journal. Nobody retypes an order and nobody cuts metal to last week's drawing.
Reference hub and protocol built 2026-09-16, 48 checks passing. Hosted phone inbox for factories with no software is in progress. Until then a factory joins through a signed HTTP call an operator runs for it.
The exchange is a plain HTTPS JSON API with an OpenAPI description. Claude, ChatGPT or your own software can read a thread and answer an RFQ with a key you hold. Custodial keys, where the hub holds the key for a factory with no software, are disclosed on every receipt and are a stopgap until the factory holds its own.
Thirty minutes on a call, then a walk of one area of your floor with the twin on a laptop. No preparation needed.