Describe features and capabilities of agents in Microsoft Copilot Studio
What an agent is
An agent built in Microsoft Copilot Studio is a conversational assistant created without code and published to channels such as Microsoft Teams, a website, or a contact-center integration. It answers questions from knowledge sources and takes actions in other systems.
The building blocks
- Topics: authored conversation paths with trigger phrases, questions, conditions, and messages. Deterministic.
- Generative answers: when no topic matches, the agent retrieves from knowledge and composes an answer. Probabilistic.
- Knowledge sources: public websites, uploaded documents, SharePoint, and Dataverse.
- Actions / tools: call a Power Automate flow, a connector, or a prompt to do real work such as creating a ticket or looking up an order.
- Orchestration: classic (topic-first) versus generative (the model selects topics, tools, and knowledge).
- Channels and authentication: publish to Teams, web, or custom channels; choose no authentication, Microsoft Entra, or manual authentication.
- Analytics: sessions, engagement, resolution rate, and unanswered questions you convert into new topics.
Traps
- Copilot Studio is where you author agents; it is not the same thing as Microsoft 365 Copilot or the Copilot pane inside Power Apps.
- Authoring changes do nothing in Teams until you publish.
- If the requirement is 'must perform X in a system of record', the answer is an action or flow, not a knowledge source.
- Knowledge sources answer questions; topics control the conversation.
Rule of thumb
Question-shaped need → knowledge source. Action-shaped need → tool or Power Automate flow. Compliance-shaped need (exact wording, regulated script) → authored topic.
Worked example: An HR agent must answer policy questions and let employees book leave. Policies = knowledge source with generative answers; leave booking = Power Automate action. Authoring a separate topic for every policy page is the wrong answer.