APL-6006 Study Guide

The quickest way to start an agent is to describe what you want it to do in plain language on the Copilot Studio Home page. Copilot Studio reads your description and generates a starting name, description, and instructions, and it suggests knowledge sources, tools, and channels you can accept or dismiss.

What you need to know

• You start by selecting the Agent tile on the Home page, or New agent on the Agents page
• The description you type can be up to 1,024 characters, and the AI turns it into the agent name, description, and instructions
• Suggestions offered during natural language creation do not persist beyond the session unless you accept them
• The natural language experience needs a supported environment and language; if it is not available, you can build the agent manually instead
• The agent name you enter before the first save sets the read-only schema name, so choose it carefully

Microsoft Learn resource: Create an agent (opens in a new tab)

Instructions are the natural language guidance that tells the agent how to behave: its role, its tone, the steps it should follow, and the rules it must not break. They are the single biggest lever you have over response quality in a generative agent.

What you need to know

• You edit instructions in the Instructions section of the agent's Overview (or Build) page
• Instructions can be up to 8,000 characters
• Type a slash (/) inside the instructions to insert a reference to a tool, topic, another agent, a variable, or a Power Fx expression
• If you created the agent with natural language, instructions are generated for you and you refine them from there
• After you change instructions on a published agent, you must publish again for users to see the change

Exam tip: Be specific in instructions about purpose, tone, and boundaries. Vague instructions are the most common cause of an agent that answers off-topic.

Microsoft Learn resource: Edit your agent's basic features (opens in a new tab)

Suggested prompts are the ready-made starter questions that appear on the agent's welcome page in Microsoft Teams and Microsoft 365 Copilot, so users know what they can ask before they type anything.

What you need to know

• You add or edit them in the Suggested prompts section of the agent's Overview page, then select Save
• You can configure up to 10 suggested prompts, each with a title and the prompt text
• They appear only in Teams and Microsoft 365 Copilot, and only before a new chat starts; you cannot see them in the Copilot Studio test pane
• Copilot Studio can generate suggested prompts automatically from the agent's description and instructions
• Publish the agent after editing suggested prompts for the change to reach users

Microsoft Learn resource: Configure suggested prompts for Teams and Microsoft 365 (opens in a new tab)

The primary language is the language the agent prefers for its responses. It is one of the identity settings you choose when you create the agent, and it matters because you cannot change it afterward.

What you need to know

• Set it in the Agent settings panel (the gear icon) before the first save, on the Agent details tab
• You cannot change the primary language after the agent is created; you can only change its region if more than one region is available
• To reach more users later, add secondary languages from the agent's Settings > Languages page instead
• For a multilingual agent, a published agent matches the user's browser or client language and falls back to the primary language when it cannot

Exam tip: Because the primary language is locked after creation, treat choosing it as a create-time decision, not something you fix later.

Microsoft Learn resource: Language support in Copilot Studio (opens in a new tab)

A solution is the Power Platform container that packages your agent and its components so you can move them between environments, for example from development to test to production. Choosing the solution at creation keeps the agent and everything you build inside it portable.

What you need to know

• Set the solution in Agent settings under Advanced settings, before you save the agent the first time
• The solution you pick determines the publisher prefix used in the agent's schema name
• Building inside a solution is what lets you export the agent and import it into another environment
• Like the schema name, these identity settings become read-only after the first save

Microsoft Learn resource: Solutions overview in Copilot Studio (opens in a new tab)

Content moderation controls how strictly the agent filters potentially harmful content in generative responses. You set the agent-wide default on the Generative AI settings page, and it applies unless a topic or prompt overrides it.

What you need to know

• The default for the agent-level setting on the Generative AI settings page is High
• A generative answers node or a prompt tool can override the agent-level default; at runtime the topic-level setting wins, and if none is set the agent-level setting applies
• Moderation levels run from Lowest to Highest on the agent's Generative AI page
• Overly strict moderation can cause the agent to return no answer at all, which is worth checking if answers go missing

Microsoft Learn resource: Knowledge sources summary: moderation (opens in a new tab)

How the agent shapes and formats its replies is driven by its generative AI settings. With generative orchestration turned on, the agent composes one formatted, cited response from the topics, tools, and knowledge it used, instead of you writing every message by hand.

What you need to know

• Generative orchestration is turned on or off in the agent's Settings > Generative AI section, under Orchestration
• With it on, the agent generates and summarizes a single response from everything it used in a turn; with it off (classic), you shape replies yourself with message nodes
• To keep a topic's wording flexible, return its result as an output variable instead of a final message node and let the agent format the reply
• For exact control over the shape of a specific answer, use the agent's instructions or an AI prompt tool, where you can set tone, constraints, and structured output such as JSON

Microsoft Learn resource: Orchestrate agent behavior with generative AI (opens in a new tab)

The model behind the agent affects its reasoning depth, response quality, and speed. Copilot Studio lets you choose the model that powers your agent's generative orchestration from a dropdown.

What you need to know

• Change the model from the agent's Overview page (or the Build tab components panel), in the Model section
• You can switch models at any time, and switching is a save-and-publish change like any other
• Experimental and preview models are for testing only, not production; they can be slower, less available, and are billed if users use them
• You can also add external models from providers such as Anthropic, Mistral, or xAI when your admin allows them
• Separate model settings exist for generative responses, deep reasoning, and the prompt builder

Microsoft Learn resource: Select a primary AI model for your agent (opens in a new tab)

The Allow ungrounded responses setting decides whether the agent may answer from the model's general knowledge, or only from your knowledge sources and tools. Turning it off keeps the agent inside your data.

What you need to know

• Find it in the Knowledge section of the agent's Generative AI settings; it requires generative orchestration
• When you turn it off, the agent blocks any response generated in a turn where it did not use a knowledge source or tool, and the fallback topic triggers instead
• With it off, the agent returns a knowledge answer only when the response includes an in-text citation to the source
• Turning it off does not guarantee the model never blends in general knowledge; it only blocks turns that used no source or tool at all

Exam tip: Disabling ungrounded responses is the setting to reach for when the goal is "answer only from our documents, nothing else."

Microsoft Learn resource: Allow ungrounded responses (opens in a new tab)

Web search, shown as Use information from the web, lets the agent pull real-time public information from Bing beyond your own knowledge sources. Turning it off keeps the agent from reaching out to the open web.

What you need to know

• The setting appears both on the Generative AI settings page and in the Knowledge section of the agent's Overview page
• It requires generative orchestration to be turned on
• When on, web results are interleaved with results from any public websites you added as knowledge sources
• It uses Grounding with Bing Search, so turning it off is the way to stop the agent using open-web content

Microsoft Learn resource: Knowledge source agent settings (opens in a new tab)

Knowledge sources ground the agent's answers in your data instead of the model's general knowledge. You can add them when you create the agent or later, at the agent level or inside a topic.

What you need to know

• Common sources include SharePoint, Dataverse, public websites, uploaded files, and Copilot connectors
• Add knowledge at the agent level on the Knowledge page, or at the topic level with a generative answers node
• For SharePoint, generative answers use only modern SharePoint pages and supported file types (Word, PowerPoint, PDF, SharePoint pages)
• Without a Microsoft 365 Copilot license in the same tenant, SharePoint files are processed up to 7 MB; with one and semantic search on, up to 200 MB; uploaded files can be up to 512 MB
• Better retrieval quality comes from a Microsoft 365 Copilot license in the same tenant as the agent

Microsoft Learn resource: Knowledge sources summary (opens in a new tab)

Generative answers let the agent search your knowledge sources and summarize a conversational, cited response, instead of you authoring a topic for every possible question. You add it as a node inside a topic, often before the Fallback topic.

What you need to know

• Add a Create generative answers node from the Advanced node menu in a topic
• Point the node at one or more knowledge sources, such as SharePoint, public websites, or a custom data source
• Set the node's content moderation level from its Properties pane; the node-level setting overrides the agent default
• When searching SharePoint, only the top few results are summarized, so aim the node at content that is relevant and indexed
• Custom data sources should include the ContentLocation (URL) and Title fields so the answer can be cited

Microsoft Learn resource: Use generative answers in a topic (opens in a new tab)

The Conversation Start topic is a system topic that controls the agent's opening message. Editing it is how you set the greeting and any first question the agent asks.

What you need to know

• Open it from Topics > System > Conversation Start
• The first Message node holds the introductory text users see; you can use variables such as the agent name inside it
• You can add nodes below the message, for example an Ask a question node to start a guided flow
• If you publish to Teams only and use suggested prompts, consider turning this topic off so the welcome page shows

Microsoft Learn resource: Use system topics (opens in a new tab)

You can have Copilot build a topic for you by describing what it should do. Copilot generates the trigger phrases and an initial set of nodes, which you then refine.

What you need to know

• Go to the Topics page and select Add a topic > Add from description with Copilot
• Give the topic a name and describe, in plain language, the questions to ask and the messages to show
• Copilot can create Message, Question, and Condition nodes; other node types you add yourself
• The generated topic includes trigger phrases based on your description
• Copilot does not set advanced node properties such as reprompt or interruption settings; you edit those manually

Microsoft Learn resource: Create and edit topics with Copilot (opens in a new tab)

You cannot delete system topics, but you can turn them off when you do not want their behavior, such as turning off Conversation Start for a Teams-only agent.

What you need to know

• System topics can be turned off but not deleted, and you cannot create new system topics
• You can edit a system topic's nodes and, for some, its trigger phrases
• Turn a topic on or off from the Topics list using the topic's on/off control
• Wait until you are comfortable building full conversations before you change system topics, since they support core behaviors like escalation and end-of-conversation

Microsoft Learn resource: Manage topics (opens in a new tab)

The Message node sends text from the agent to the user. It is the simplest node and the one you use most, and it can carry more than plain text.

What you need to know

• Add it with the Add node (+) icon, then Send a message
• Beyond text, a message can include variables, message variations, images, videos, basic cards, Adaptive Cards, and quick replies
• Insert a variable with the {x} icon in the node's menu bar
• You can rename nodes to make a topic readable; node names can be up to 500 characters

Microsoft Learn resource: Send a message (opens in a new tab)

The Question node asks the user something and stores the answer in a variable. What you choose under Identify decides how the agent interprets the reply.

What you need to know

• Add it with Add node (+) > Ask a question, then enter the question text
• Under Identify, pick how to read the response, for example Multiple choice options or a specific entity
• For multiple choice, each option you enter becomes a button and Copilot Studio creates a conditional path for it automatically
• The answer is saved in a variable you can rename and reuse later in the topic
• Set a variable to Receive values from other topics so a redirect can fill it and skip asking again

Microsoft Learn resource: Ask a question (opens in a new tab)

A Redirect node hands the conversation from one topic to another without starting over, which lets you build small reusable topics and call them from several places.

What you need to know

• Add it with Add node (+) > Topic management > Go to another topic, then pick the destination
• By default the conversation returns to the original topic when the redirected topic ends, so you can add more nodes after the redirect
• You can pass input variables to the destination topic and receive output variables back
• Redirecting to certain system topics (End of Conversation, Goodbye, Escalate, Start over, Confirmed Success, Confirmed Failure) ends the whole conversation instead of returning

Microsoft Learn resource: Redirect to another topic (opens in a new tab)

A Condition node branches the conversation based on the value of a variable, so the agent can respond differently depending on what it knows.

What you need to know

• Add it with Add node (+) > Add a condition; conditions need variables to be available in the topic
• Select Branch based on a condition to add a path for a specific value
• Select Branch for all other conditions to add the catch-all path for every other value
• A common use is showing different messages for a signed-in versus an anonymous user

Microsoft Learn resource: Use conditions to create branching logic (opens in a new tab)

An Adaptive Card shows an interactive card with buttons or input fields, which is useful when a plain message or a simple question is not rich enough.

What you need to know

• Add it as its own node, or from the tools inside a Message or Question node
• Cards can present one or more response buttons and input fields to collect several values at once
• If you render a generative answer yourself through a card instead of a Message node, citations are not added automatically; you handle citation rendering
• Adaptive Cards with mixed static text and variables need a workaround to localize, so plan for that in multilingual agents

Microsoft Learn resource: Ask with an Adaptive Card (opens in a new tab)

Adding a tool at the agent level makes it available to generative orchestration, so the agent can decide on its own when to call the tool based on the tool's name and description.

What you need to know

• Add it from the agent's Tools page with Add a tool, or from the shared Tools page with Add to agent
• Tool types include prompt, agent flow, connector, custom connector, REST API, Model Context Protocol, and computer use
• When you add an existing tool to an agent, Copilot Studio wraps it so you can set inputs, outputs, and behavior specific to that agent
• With generative orchestration on, a well-described agent-level tool is called automatically when user intent matches; you do not have to wire it into a topic

Exam tip: Add a tool at the agent level when you want the orchestrator to choose it by intent; add it inside a topic when you need deterministic, scripted control.

Microsoft Learn resource: Add tools to custom agents (opens in a new tab)

Adding a tool inside a topic gives you deterministic control: the tool runs at the exact point in the conversation where you place the node, in the order you decide.

What you need to know

• Open a topic, select Add node (+), choose Add a tool, then pick a connector, flow, or other tool
• You may be asked to create or select a connection for the tool
• By default a tool connection uses the end user's credentials; you can change the authentication mode
• Use a topic-level tool for scripted flows with an exact order of steps and validations, rather than leaving the choice to the orchestrator

Microsoft Learn resource: Add a tool from a prebuilt connector in a topic (opens in a new tab)

Inputs are the values a tool needs to run. In a generative agent, the agent can gather them for you, or you can override an input with a fixed value.

What you need to know

• Configure inputs on the tool's configuration page, in the Inputs section
• By default each input is set to Dynamically fill with AI: the agent pulls the value from context or asks the user for it
• Set Fill using to Custom value to override an input with a literal value, a variable, or a Power Fx formula; the agent then does not ask for it
• Identify as sets how the user's answer is interpreted (a string or a specific entity), and you can add retry logic and extra validation

Microsoft Learn resource: Configure your tool inputs (opens in a new tab)

Outputs are the values a tool returns and how the agent responds once the tool finishes. You choose which outputs are available and what the user sees.

What you need to know

• Configure what happens next in the tool's Completion section, under After running
• Options include Don't respond (the agent folds the output into its own answer), Write the response with generative AI, Send specific response, and Send an adaptive card
• A specific response can insert output variables with the variable picker and format them with Power Fx
• You can choose which output variables to expose so other tools and the agent can reuse them

Microsoft Learn resource: Configure tool completion and outputs (opens in a new tab)

Naming a tool in the agent's instructions tells the orchestrator when and how to use it, which makes tool selection more reliable than leaving it to the description alone.

What you need to know

• Edit instructions in the Instructions section, and type a slash (/) to insert a reference to the tool
• A clear tool name and description are what the orchestrator reads to decide when to call the tool, so keep them specific
• Referencing the tool in instructions is how you guide the agent to use it for the right task at the right time
• Publish the agent after editing instructions so the change is live

Microsoft Learn resource: Write agent instructions (opens in a new tab)

Sharing grants other people access to the agent so they can use it, or in some cases test or co-author it. Because an agent must be published before it can be shared, Copilot Studio prompts you to publish first if you have not. You can share with individual users, security groups, or your whole organization.

What you need to know

• Publish first: select Publish and confirm (a green banner shows success), then turn on channels such as Teams + Microsoft 365, and Make agent available in Microsoft 365 Copilot to surface it there
• Open the agent and select the Share icon near Publish; if the agent is not published you are prompted to publish first
• Enter users, email addresses, or security groups, then select Share; people you add start with a Pending status
• Set organization access to End user access so anyone in your org can find and use the agent, or restrict it to only the people you specify
• Users you share with generally need a Copilot Studio license and Microsoft authentication (Authenticate with Microsoft)
• Viewing or editing rights depend on the experience and on Dataverse security roles such as Environment Maker

Exam tip: You cannot share an agent that has not been published, so publish before you share.

Microsoft Learn resource: Share agents with other users (opens in a new tab)