One agent can do one job. A team of agents needs a supervisor: the agent that reads every request and hands it to the right specialist. In Agent Studio this is called a supervisor team, and it is how one chat window can serve order lookups, delivery tracking, procedure answers, gated changes and stuck approvals at once. You are the supervisor: 8 real requests from a supply chain floor. Route each one to the right specialist. Every miss teaches you a routing rule.
Checked Applies to Fusion Cloud 26C · Routing and Switch node facts verified 19 Jul 2026 against the AI Agent Studio guide (26C).
An original training scenario, checked against the current release. Multi-agent capabilities in Fusion change every quarterly update, verify against the current release. Companions: the AI Agent Studio Explained · Which Agent Type to Use.
The Switch node is one of the Workflow Control nodes (alongside If Condition, For Loop, While Loop, Parallel, and Wait). It evaluates a runtime value and routes the workflow to the branch that matches. The common routing pattern pairs it with an LLM node: the LLM node reads the user request and extracts the intent as a value, and the Switch node sends each intent down its own branch. Name every branch after the intent it serves. When a request goes down the wrong branch, open the run results: they show the value the Switch received and the branch it chose, and the canvas highlights the path taken.
A supervisor team routes with an LLM: the supervisor reads the request and the descriptions of its agents and tools, then decides. That is the right choice for open conversations where the path cannot be fixed in advance, and it is exactly what you practiced above. A Switch node routes deterministically: the same input always takes the same path, which auditors and support teams love. The pattern that scales in production combines them: one entry point, an LLM node that classifies the request into a small fixed list of intents, a Switch node that sends each intent to its own specialist agent, and an intent value reserved for everything the classifier cannot place, routed to a branch that asks a person instead of guessing.
Three reasons that show up on the invoice, not just the diagram. Accuracy: each specialist keeps short, focused instructions and only the tools its job needs, so it stops confusing itself. Speed and cost: a routed status lookup skips the reasoning step a do-everything agent pays on every request. Model choice: each specialist can run a different model, and a receipt-status lookup does not need the model that interprets policy documents. Example from our own domain: a warehouse assistant with one entry point, where the Switch routes receipt-status questions to one specialist, item-availability questions to another, and transfer-order requests to a third that ends in a Human Approval node before anything is created.