Choosing the right agent approach in Copilot Studio
Most failed agent projects pick the wrong building block. Before you open Microsoft Copilot Studio (formerly Power Virtual Agents), decide what the agent must do, who starts it and how predictable it must be.
Conversational or autonomous
| Conversational agent | Autonomous agent | |
|---|---|---|
| Started by | A user message | An event trigger or schedule |
| Typical use | Q and A, guided help | Process a new email, row or file |
| User in the loop | Yes | Usually no |
An autonomous agent still uses instructions, knowledge and tools. The trigger replaces the chat message, and its payload becomes input for the agent. Plan which identity runs the agent and what data it may touch, because nobody is there to approve each step.
Orchestration mode
- Generative orchestration (the default for new agents): the model reads the request and picks topics, tools, knowledge sources or other agents. It relies on good names and descriptions.
- Classic orchestration: topics fire only when a trigger phrase matches.
Use classic when behaviour must be fully scripted. Use generative when users ask open-ended questions or one request needs several steps.
Building blocks
- Instructions: overall behaviour, tone and limits.
- Knowledge sources: content to ground answers (SharePoint, websites, files, Dataverse, Copilot connectors).
- Tools: connectors, prompts, agent flows, REST APIs, MCP servers, computer use.
- Topics: scripted conversations for cases that need control.
- Agents: child or connected agents for splitting a large job.
Agent flows
An agent flow is a Power Automate flow built for agents. Use it when steps must run exactly the same way every time, such as create a record, then request an approval, then update a status. The agent decides when to call it, but the flow itself is deterministic.
Planning checklist
- List the tasks and mark each as open-ended (generative) or fixed (flow or topic).
- Identify data sources and who may see them.
- Pick channels and authentication.
- Decide which environment holds the agent and which DLP policies apply.
- Define success measures you can see in analytics.
Exam traps
- Autonomous does not mean no instructions. It still needs instructions, tools and a defined trigger.
- A topic is not required for every task. Generative orchestration can call a tool directly.
- Agent flows are deterministic. Do not use them for open-ended reasoning.
- Poor descriptions on tools and topics are the usual reason the wrong one is chosen.
- Power Virtual Agents is the old product name. The exam may still use it.
Practice questions
Choose an answer, then review the explanation to check your understanding.
1. An agent must start work by itself when a new row is added to a Dataverse table, with nobody typing in chat. What do you add?
Show answer
Answer: B.Event triggers (for example a Dataverse or SharePoint change) let an agent act autonomously without a user message.
2. A team wants a deterministic, step-by-step process with fixed logic that the agent can call as one action. What should they build?
Show answer
Answer: A.An agent flow runs fixed, predictable steps and can be called from the agent as a tool.
3. Which choice describes generative orchestration?
Show answer
Answer: B.With generative orchestration the model chooses which topic, tool or knowledge source to use from their names and descriptions.