3.2.1. Creating Agent Flows
💡 First Principle: A flow is a trigger plus a pipeline of actions — and in Copilot Studio, the trigger is the design decision. "When an agent calls the flow" makes it a tool the orchestrator or a topic can invoke; a schedule or event makes it autonomous automation that runs with no conversation at all.
Creation offers two on-ramps. Describe the automation in natural language and Copilot Studio drafts the flow — triggers, actions, connections — for you to refine; or build in the designer directly, assembling actions, conditions, and loops with AI-assisted suggestions. Either way the action vocabulary is the same: connector actions (the Phase 2 ecosystem again), built-in logic (branching, looping, data operations, child flows), AI capabilities (prompts, document processing), and human-in-the-loop steps (3.2.2).
Billing awareness is now part of authoring: flow runs consume Copilot Studio capacity per action (testing in the designer is free), and when prepaid capacity runs out, new runs are blocked — which makes capacity planning a reliability concern, not just a cost line. An always-running scheduled flow and a chat-triggered flow draw from the same envelope.
⚠️ Exam Trap: For a flow an agent should invoke as a tool, the trigger must be "When an agent calls the flow." A scheduled or manually triggered flow won't appear as an invocable tool — trigger choice is the first thing to check when "the agent can't see my flow."
Reflection Question: You need (a) a nightly sync of ticket statuses and (b) an on-demand "create a return label" capability in conversation. Same actions library, different triggers — which trigger each, and why does mixing them up fail structurally?
