1.1.2. Generative Orchestration in Copilot Studio
💡 First Principle: Generative orchestration inverts control. In classic orchestration, your trigger phrases decide which topic runs — the author is the router. In generative orchestration, the model is the router: it reads the user's goal plus the names and descriptions of every topic, tool, and knowledge source you've attached, and plans which ones to use, in what order, filling required inputs as it goes.
This is why descriptions suddenly matter so much. Under classic orchestration, a topic's trigger phrases were its doorway; under generative orchestration, its description is. The orchestrator can only choose well among components it can understand — a tool described as "Utility flow 3" is invisible to good planning, while "Creates a purchase order in SAP given a vendor ID and amount" practically routes itself. On the exam, "the orchestrator isn't selecting my tool" scenarios are usually description problems, not code problems.
Generative orchestration also enables behavior classic mode cannot deliver: chaining multiple tools in one turn, asking the user for a missing required input without you authoring a question node, and falling back to knowledge when no tool fits. Classic orchestration remains available — and remains the right answer when a regulated process demands that specific utterances always follow one exact authored path.
⚠️ Common Misconception: Generative orchestration means the agent improvises everything and topics no longer run. In reality, the orchestrator selects among your topics, tools, and knowledge — but an authored topic still executes its nodes deterministically once selected. You keep deterministic execution; you delegate selection.
⚠️ Exam Trap: Switching an existing agent from classic to generative orchestration changes how topics are chosen (descriptions instead of trigger phrases). A topic that worked via trigger phrases may stop firing until its description accurately explains what it does.
Reflection Question: Why would a bank's "report fraud" flow be a candidate for classic orchestration — or for a generative agent, a tightly-described topic — while "answer questions about our credit cards" fits generative orchestration with knowledge?
