3.2.2. Agent Flows and Orchestration
Agent flows are deterministic, rule-based automation built in Copilot Studio: the same input always produces the same output. A flow with the When an agent calls the flow trigger can be added to an agent as a tool. The conversation logic — routing, questions, messages — lives in topics and the agent's generative orchestration; an agent flow holds the predictable sequence of steps (connector calls, AI actions, approvals, loops and branches) that runs when the agent calls it.
Agent flows are NOT the same as Power Automate flows, and the exam tests this distinction:
In short: agent flows are authored inside Copilot Studio and exposed to the agent as deterministic tools it can invoke and get results back from, while Power Automate cloud flows are standalone automations in the broader Power Automate service that run independently of any agent.
| Dimension | Agent Flows (Copilot Studio) | Power Automate Flows |
|---|---|---|
| Purpose | Deterministic, rule-based task automation that an agent calls as a tool | Workflow automation and system integration |
| Execution | Runs when the agent calls it (or on a schedule or event) and returns its result to the agent | Background execution, often asynchronous |
| Triggers | Called by an agent (from a topic or through generative orchestration); instant, scheduled, or event triggers | System events, scheduled, HTTP requests |
| Actions | Connector actions, AI actions (such as running a prompt), approvals and other human-in-the-loop steps, loops and branches | Data operations, approvals, file management, 1000+ connectors |
| When to use | The agent needs a predictable, repeatable action sequence it can call as a tool | Process needs to run independently of the conversation |
Agent Flow Design Patterns:
Linear flow — Steps execute in sequence. Receive inputs from the agent → validate → process → return the result. Best for straightforward task completion.
Branching flow — Decision nodes route to different paths based on conditions. "Is the order domestic or international?" leads to different processing logic. Best for tasks with conditional logic.
Looping flow — Repeat a sequence until a condition is met. Iterate through a list of items or repeat a step until it succeeds. Best for batch and iterative processing.
Composite flow — Combines the conversation with flow calls. The agent handles the conversation, calls an agent flow for the backend operation, then continues the conversation with the results the flow returns. This is the most common enterprise pattern.
Design Best Practices:
- Keep conversation logic in topics and the agent's orchestration, and keep agent flows focused on the deterministic task. Collect inputs in the conversation, pass them to the flow, and let the flow run the data operations, calculations, approvals, and system integrations.
- Handle errors explicitly. If an agent flow call fails or times out, the agent should inform the user and offer alternatives — not silently fail.
- Minimize round-trips. Each call from an agent flow to an external system adds latency. Batch related data requests where possible.
- Don't block on slow backends. If a backend step is slow (for example, an underwriting API that takes 10-15 seconds), have the agent collect the details in the conversation and start the backend process asynchronously (through an agent flow), tell the user it is underway, and report back when it completes — rather than a long synchronous wait, a progress bar, or skipping the validation.
Reflection Question: An agent needs to help a user book a conference room: check availability, confirm details, send invitations, and update the calendar. Which parts belong in the conversation (topics) vs. in an agent flow the agent calls?