7.1 Building Agents: Topics, Knowledge, Tools (MCP, Agent Flows), Publishing

What Is Microsoft Copilot Studio?

Microsoft Copilot Studio is the low-code tool within the Power Platform for building AI-powered agents — conversational assistants that can answer questions, complete tasks, and take action on a user's behalf. On the PL-900 exam, "agent" is the current term for what earlier materials called a bot or copilot: an AI-driven helper a maker builds without writing code, then deploys into the places people already work, such as Microsoft Teams or a company website.

An agent in Copilot Studio is assembled from four core building blocks: topics, knowledge, tools, and the channels it publishes to. Understanding what each block does — and when a maker would reach for it — is the heart of this exam objective.

Topics: Defining Structured Conversation Paths

A topic is a single, structured conversation path inside an agent. Each topic starts with one or more trigger phrases — sample user utterances such as "reset my password" or "check my order status" — that tell the agent's natural-language understanding when to enter that path. Once triggered, a topic's dialog walks the user through a defined sequence: asking a question, sending a message, evaluating a condition, or calling an action such as a flow.

Copilot Studio ships every new agent with system topics already in place, including:

• Greeting — the agent's opening message
• Escalate — hands the conversation to a human agent
• Fallback — triggers when the agent cannot match user input to any topic or knowledge source
• Multiple Topics Matched — asks the user to clarify when more than one topic could apply

Makers add custom topics on top of these system topics for scenarios specific to their business, such as an HR agent's "request time off" topic. Because trigger phrases only need a handful of examples, Copilot Studio's underlying AI generalizes to recognize similarly worded requests, reducing the amount of dialog a maker must author by hand.

Knowledge Sources: Grounding Generative Answers

Not every question a user asks can be — or should be — mapped to a hand-built topic. Knowledge sources let an agent generate accurate answers directly from existing content instead of relying on dialog a maker wrote in advance. Copilot Studio can ground its generative answers in:

When a user asks a question that doesn't match any topic's trigger phrases, the agent's generative AI orchestration searches connected knowledge sources, retrieves relevant passages, and composes a natural-language answer — with no dialog authored by the maker. This is what allows an agent to competently field the long tail of questions a maker could never anticipate one topic at a time.

Tools: Giving Agents the Ability to Act

Tools extend an agent beyond answering questions so it can take action — updating a record, sending an email, or looking up live data. Copilot Studio agents can call several kinds of tools, including prebuilt connectors to Microsoft and third-party services and AI prompts, plus two capabilities the PL-900 blueprint calls out specifically:

• MCP servers — Model Context Protocol (MCP) is a standardized, open way for an AI agent to discover and call external tools and data sources. Because MCP is a common protocol rather than a proprietary connector, an agent can connect to any MCP-compliant server to reach outside systems consistently, without a maker building a custom integration for every source.
• Agent flows — a Power Automate-style flow designer purpose-built to run as a tool inside an agent. Rather than being triggered independently the way a traditional cloud flow is, an agent flow is a deterministic sequence of steps the agent itself decides to invoke mid-conversation — for example, to look up an order or submit a request — then uses the result to keep responding to the user.

Together, tools are what separate a simple question-answering chatbot from an agent capable of completing real work on a user's behalf. A single agent can combine several tools — for example, an HR agent might use a connector to look up an employee record, an agent flow to submit a time-off request, and an MCP server to query a third-party benefits system.

Generative AI Answers vs. Hand-Built Dialog

Keep two response styles straight for the exam. A topic's dialog is authored by a maker node by node, so its wording and logic are entirely predictable. A generative AI answer grounded in a knowledge source, by contrast, is composed by the underlying model at run time from whatever content it retrieves — the maker never writes that exact sentence. Most production agents blend both: topics handle high-value, must-get-right conversations, while knowledge-grounded generative answers absorb everything else a maker didn't script in advance.

Publishing Agents to Channels

Building and testing an agent happens inside Copilot Studio's authoring canvas and test pane, but none of that work reaches real users until the maker publishes it. Publishing is a distinct step from saving changes: saving commits edits to the agent's definition, while publishing pushes the current published version live to every channel the agent is connected to.

Copilot Studio agents can be published to multiple channels at once, including:

• Microsoft Teams, where users chat with the agent like any other Teams contact
• A custom website, using an embeddable web chat control
• Other messaging and custom channels supported through the platform's channel connections

Because publishing is separate from saving, a maker can continue editing and testing an agent in the authoring canvas without affecting the version end users currently interact with — changes only go live the next time the maker publishes them.