Getting Started with Microsoft Copilot Agent Builder 🚀 | Build Your First AI Agent | ShareTech Live

Vídeo em Inglês, publicado pelo canal ShareTech.

About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.

Video summary

The video demonstrates how to use Microsoft Copilot Agent Builder to create, configure, test, and refine a custom AI agent called Microsoft Learn Assistant. The presenter covers defining agent purposes, setting instructions, integrating official Microsoft documentation as knowledge sources, testing prompts for technical questions and troubleshooting, and incorporating reusable skills.

Key points

• Introduction to Microsoft Copilot Agent Builder within the M365 ecosystem and its low-code experience for building purpose-specific AI agents.
• Explanation of the core building blocks of an agent: purpose, instructions, knowledge sources, capabilities, and testing.
• Creation of a 'Microsoft Learn Assistant' using a natural language prompt to cover documentation on technologies like SharePoint, Azure, and Microsoft Graph.
• Configuration of official Microsoft Learn documentation URLs as specific knowledge sources to ground the agent's responses.
• Testing the agent with practical queries, such as SharePoint site architecture, creating SPFx web parts, and handling troubleshooting scenarios without inventing information.
• Demonstration of adding modular skills via a zip file containing a skill.md file to extend the agent's repeatable tasks.

Topics

Microsoft CopilotAgent BuilderAI AgentsKnowledge SourcesMicrosoft LearnSPFx DevelopmentTechnical TroubleshootingAgent Skills

Video transcript

Transcript from available captions; it may contain errors.

Hello everyone and welcome to Share Tech's Agent Live with me session. I'm Rahul Trimukhe and today just give me a second. Hey hi. Okay, I'm Rahul Trimukhe and today we are going to explore something exciting in the Microsoft ecosystem. Microsoft Copilot Agent Builder. Hope you're able to see my screen. I have already opened M365 agent store where we can see the agents created by me, the agents created by Microsoft. And yeah, thank you for taking the time to join whether you are a developer, a Microsoft administrator, a business user, or someone who is simply curious about AI. This session is designed to help you understand how you can start building your own AI agents. Before we begin, one small request, please feel free to ask questions throughout the session. You can post your questions in the comments or chat if you're joining live and uh would you like to discuss something, you're welcome to unmute yourself and ask your questions. You don't have to wait until the end if something isn't clear, just ask it. Okay. So today I won't I I won't just explain what an AI agent is, we will actually build one together, configure its instructions, connect knowledge sources, and test how it responds. By the end, I want you feel confident enough to open Agent Builder and start creating an agent for your own use case. So, let's get started. Before we start building, let's understand what we are actually building. We all use AI assistance to ask questions, summarize information, draft emails, and generate ideas. But, what if we want an AI assistant that has a specific role, follow our our own instructions, and use a selected knowledge to help us with a particular task? That's where agent comes in. For example, imagine a new employee joining your organization. Instead of asking different people about leave policies, onboarding procedures, uh internal tools, and company processes, they could ask an onboarding agent that uses the approved information you have provided. Or, if you're working with SharePoint, you might uh build an assistant that answers questions about your team's documentation. The important thing to understand is that an agent is not simply a name attached to a chatbot. We define its purpose instructions knowledge sources, and where supported, and their capabilities, too. Microsoft Copilot Agent Builder helps us do this using a guided low-code experience. We will start with a simple use case and build it step-by-step. For today's uh demonstration, I'm going to create a shared uh a Microsoft Learn Assistant. Just give me a second. >> Allow me a minute. In case if you have question, you can just post it in comments before getting started. Okay, we will start by understanding what an agent is. Uh then we will open Microsoft Copilot Agent Builder and create our own Microsoft Learn Assistant, configure its instructions, add official Microsoft documentation as a knowledge source, and test its answers. We will also discuss skills and understand how they can extend an agent beyond a basic uh question and answer experience. Our example today uh will be a Microsoft Learn Assistant. So, imagine you are learning you are learning SharePoint, Azure, Microsoft Graph, or another Microsoft technology instead of manually navigating documentation for every question. you want an assistant that can help you understand the concept and guide you through the documented steps. That's the scenario we are going to explore today. Let's open Microsoft 365 Copilot Studio to get started. I have already opened it in my uh browser and hope you guys are able to see it. And let's start with a simple question, what exactly is an AI agent? As I have already mentioned you, uh an AI agent is an AI-powered assistant configured for a particular purpose. It can follow instructions, use available knowledge, and depending on its capabilities, perform certain task. Think of generally AI assistant as a someone you can ask about anything. An agent is more purpose-specific. We define what it should help and uh how it should behave. So, when we build an agent, there are few important building blocks. First, we have instructions. Uh this tells the agent its role and how to respond. Second, we have knowledge. This provides relevant information the agent can use. Third, we have its capabilities. These determines what additional task it can perform. Finally, we have testing. We need to validate that the agent behaves as expected. So, remember the simple formula, purpose instructions knowledge and testing. Purpose, what problem should the agent solve? Instruction, how should the agent behave? Knowledge, which information should it refer? And the testing, does it respond reliably and expected output we are getting it. Okay? So, Microsoft Copilot Agent Builder gives us a guided way to create declarative agents for a specific scenario in Microsoft 365 Copilot. We can describe an agent in a natural language, configure its instructions and knowledge, and test it before using or sharing it. We don't need to begin by writing code. That makes it a useful starting point for people who want to understand agent creation before moving into the advanced development. For today's demo, we are going to create an agent called Microsoft Learn Assistant. Our goal is to help users understand Microsoft technologies through official documentation. We will focus on practical questions, explain explaining its concepts, understanding development procedures, and troubleshooting technical issues. Issues. Let's move from the concept to the actual build. Here we are in the agent creation experience. We have the option to describe what uh So yes. So, here we have an option to describe what we want in the natural language. Agent Builder can use that description to propose an initial configuration. This is useful because we can focus on defining the problem before worrying about every configuration details. I have a prompt already created. I'll just copy that and paste it in this input box. Okay, just to read out this prompt for you. Create an agent named Microsoft Learn Assistant. Help developers, IT professionals administrators and learners understand Microsoft technologies using official Microsoft Learn documentation. Which covers Microsoft 365 Copilot, SharePoint SPFx, Azure, Microsoft Entra, Power Platform, and Microsoft Graph. Explain concepts, summarize documentation, provide step-by-step guidance, and help troubleshooting technical issues. These are the capabilities. Prefer configured Microsoft Learn knowledge source for Microsoft specific questions. Never invent documentation links, product features, article titles, or technical requirements. If the source do not contain enough information, say directly and say so clearly. Use a friendly tone and professional tone. This is what the instruction is I have written. I'm describing the purpose of the agent, the audience it serves, the technical topics it covers, and how it should handle missing information. Notice that we are not giving it every possible answer. We are defining its behavior and then we will supply documentation as a knowledge source. This separation is important. Instructions guide the agent, knowledge provides the reference material. Okay, let me send this prompt now. Agent builder has started the generation of the initial state up. Now we are going to review it once it has been created. And here we go. I want the name to be clear, the purpose to be specific, and the behavior to be appropriate for technical learning. We will refine the configuration and then add Microsoft Learn as a knowledge source. So, here we can see the name of the agent description and the instructions are also generated by the agent builder. Below to that, we have an option to add skills. We will see the skills section at the end of the session. And then we have an option to add knowledge source where we're going to add Microsoft Learn site as a knowledge source to this agent. Okay? And this is the place where we can inspect the agent settings rather than relying entirely on the generated descriptions. We'll We'll be able to like review its name, description, instructions, knowledge sources, and its capabilities. Everything will be available in this configuration pen. Okay. So, what I'm going to do I don't I'll just replace these instructions generated by the bot, the one which I have created for my agent, okay? And then I'll save it. So, these instructions tells the agent how to answer technical questions, when to ask for clarification, and how to handle unsupported information. The goal is not to make the instructions unnecessarily long. The goal is to make them clear enough that we can test the behavior consistently. And we're going to add the knowledge source now. Just give me a second. Here we have a new button called add a knowledge source where we will specify this uh URL for the SharePoint knowledge source. I'll just open this in new tab. What it is. Okay, this is the documentations. Which we have for SharePoint. Okay. So this is focused starting point for SharePoint documentation. For this first built uh use one source and verify that it works before adding another. Now we are adding the knowledge source. For this documentation, we are starting the official Microsoft learn SharePoint documentation. This keeps our first test focused and we will expand the source coverage if we need more topics. Microsoft learn covers a huge range of technologies, but adding one documentation URL doesn't automatically mean every page across Microsoft learn is available to the agent. The configured URL scope matters. We should use a supported URL. Verify it it appears in the configuration and test whether the agent can retrieve the information we need. If you want to cover multiple product families, we can add other focused Microsoft learn URLs within the supported limits. Microsoft currently documents a maximum of four public website URLs per agent with the URL depth restrictions. So adding a source is not the same as proving the retrieval works. We will verify that the by the asking questions whose answer should be available in the documentation. That is one of the most important checks in agent development. Okay, let me now click on create and then the agent will get created. Okay. So, the agent is now created. We have a preview section here where we can execute the prompts. I'll just select this one and send. We'll see whether we will be able to retrieve the response or not. So, this is what our agent has generated the information. And if you see the URL here, which has been referred by the knowledge source which we have added. Okay, so it seems it's referring to the correct knowledge source. Let me just configure it again and add all other remaining knowledge source to it. Let's add for Azure and M365. Okay. We can extend the agent to cover the Microsoft technologies by adding more focused documentation sources. I'm keeping our initial scope manageable so we can validate the experience before making it uh broader. So, learn SharePoint explains site architecture and its main components. Understand SPFX, explain SharePoint framework and how it is used to build web parts. Follow a procedure. Give me Give me the documented steps to create SharePoint framework web part. So, these are the user prompts which we going to add. Learn SharePoint. Okay. I want to keep that on top. One more. Understand SPFX. And then procedure. Keep this here, okay? So, suggested prompts make the agent easier to use. They give users starting point and help communicating Just give me a second. What kind of questions the agent is intended to answer? So, the capabilities extend the experience beyond basic question and answer. For example, code interpreter can suggest can support certain data analysis and code related task while image generation and uh image generation can create images from the prompts, but our our core use case today is technical learning. We don't need to enable every available feature to make a useful agent. Keep optional capabilities disabled unless they contribute to your uh demonstration. Uh like your practical use case. Okay? So, these are the capabilities like create images, create documentation chart code discover model preferences. And here in the six settings, you'll be able to find uh rest of the other options. Earlier, that capability option were visible like here itself. Now, it's not there. Okay? So, we have defined the purpose, added instructions, configured our agent uh knowledge source, and selected the our available settings. Let's create the agent and see how it performs. Let me update. After creation, we will test it rather than assuming that the configuration is automatically correct. Okay? So, what we going to do is like we will be saying I'm starting with the broader question uh to see whether the agent can explain the basic uh topic clearly. I'm looking for a structured explanation that is useful to a beginner. I'll uh also check whether the response provided by the agent is relevant and from the accurate document references. So, SharePoint site architecture overview. He has explained what is tenant site libraries, pages, and permissions, hub sites. Okay, most of the details seems to be an accurate one. What are the pages, permissions, and security, security groups, search, content types, and metadata, integration layer, modern SharePoint architecture. Uh which contains Microsoft tenant hub sites inside that associated hub sites, multiple hub sites with their associated uh associated sites. Okay. And then key component summaries mentioned were here. Amazing. So yeah. It seems the data is getting generated correctly. Uh let me ask another question. This one. So, based on Microsoft Learn, explain how to create a SharePoint Framework web part, include the prerequisites and main steps. Now, we are testing a more practical question. Instead of just asking for a definition, we are asking for a documented procedure. This helps us evaluate whether the agent can organize technical information into a useful learning path. Uh okay, these are the prerequisites, skills and access development environment, which needs Node.js, Visual Studio SPFx Yeoman is the scaffolding generator in which we will be using this SharePoint generator, modern browser, and PM. These are the commands which we need to execute to get installed Yeoman and SharePoint generator, and then create a document library uh a library on your local machine, and then run this command to go into that folder, and then install uh create the SPFx scaffolding, and then uh select the configuration like which type of component that you're going to build, name of the component, and the framework which you're going to use. Uh and then understand the project structure generated by the Yeoman. And then web part code, build configuration, package definition. So, this is the command to run. And this URL needs to be updated with your tenant URL. So, yes, to bundle use this command package use this command ship is being used, then it will generate SP package file which needs to get deployed in your SharePoint app catalog. So, typically the development workflows is set up environment, create SPFx project, generate web part, develop code, test in local workbench, test in hosted workbench, and then package solution, deploy to app catalog, add a web part to your modern page. This is the life cycle of the SPFx web part development and deployment. Amazing. So, key Microsoft learn takeaways also mentioned official references from where this information is collected is also captured by our agent. So, we're going to ask the next question here to our agent. Which Microsoft learn source are you using for this answer? A detailed uh reference we wanted to know. So, these are the four references it's referring. Even in Microsoft learn base question explanatory, I can walk through those articles sections by section. So yeah. Okay, this is where testing becomes important. We don't just evaluate whether an answer sounds convincing. We evaluate whether it's relevant, whether the technical steps make sense, and whether the documentation supports the response. If something is missing, we can refine the instructions, improve our source selection, and run the same test again. Uh I'm going Uh I'm getting an SPFx Suppose uh Suppose I'm getting an SPFx build error, and I want to troubleshoot that. So, that kind of issues also our agent should be able to resolve like will help us to get it resolved. So, I'm saying here I'm getting an SPFx build error, help me troubleshoot it step by step, and ask me for the details you need before fixing it. Absolutely. Full error message is asking the SPFx version I'm using, the node version with the command which I need to execute to get the node version. The command which gives us uh which allow us to solve the SPFx application, uh to build the SPFx application. So, that depends on like how you're using it. You're using gulp, or like you're using heft, a new technology introduced by the Microsoft. Technically, troubleshooting requires context. The right solution may depends on the SPFx version, Node.js version, package dependencies, and the exact error message. A useful agent should recognize when it needs more information instead of guessing at a fix. Okay? So, let me ask this question. What was the exact SPFx build error on my computer yesterday? This is deliberate test. We haven't provided the agent with that information. We wanted to see whether it acknowledges the limitation instead of inventing an error message or pretending to know something it has never been given to the agent. Okay. So, in case if your agent giving this unnecessary information, what you can do is let's go here to the agents tab, learn assistant, edit this. What we can do is like we can add an additional instructions okay? Here. So, go in a view mode. Below the boundaries, we can mention that for technical troubleshooting, ask the exact error message, relevant product versions, and production reproduction steps when needed. Do not assume details that the user has not provided. Prefer relevant configured Microsoft learn documentation and distinguish verified documentation for general troubleshooting suggestions. Never invent source links, okay? Fine. Let's update this agent now. After changing instructions, we should repeat the original test. This lets us see whether the refinement improvement refinement improved the behavior. Repeatable test is much more useful than the simply asking another questions and hoping for a better response. Even with good instructions and official documentation, an agent can make mistakes. A configured source may not contain the exact information needed and a generated response still needs validation. For production use, we should test important scenarios, review the source references, consider licensing and access permissions, and avoid relying on unverified answers for critical decisions. We have now completed the core agent building journey, create, configure, connect knowledge, test, and refine. Next, let's look at skills and understand how they fit into the bigger picture. So, before we talk about skills, let's make one distinction. Instructions tells an agent how to behave. Knowledge gives it information to reference. A skill packages a repeatable task with own instructions and where supported resource files or scripts. Okay? So, it's a capability, well-structured capability we can say which can be integrated with your agent. So, that is the skills. Think of a skill as a reusable procedure rather than simply another document to search. For example, our Microsoft Learn assistant answers technical questions. Uh just give me a Okay. Uh our Microsoft Learn assistant answers technical questions a separate reusable skill might take a list of changes and format them into a consistent release notes. So, yeah, for the demonstration part, what we're going to do is here we are in the configuration pane and we have a section over here, uh which is called skills. I have already created a skill. Let me show it to you. Just give me a second. Let me open a VS code, a new window, and open stop skills. Okay. See, we won't be able to upload the direct skill.md file. We need to zip it and then get it uploaded. Uh Okay. So, this is the zip file which I have already created in which we have the skill.md uh file. Let me expand this and I'll show you how it is constructed. What it's mentioned. Okay. Let me close this for now. So, there should be a name of the skill mentioned and the description what it is and for what purpose it is and the detailed purpose of the skill. It is the same way we write the instructions. Just you will just keep it separate with their purpose. Okay, keep it limited. Repeatable skills we can create and that skill we going to upload it here. Expand this. And add a skill. Stop and then skills and then I'll choose this one. Then upload. Okay, the skill is now uploaded and all the instructions you'll be able to see. And in case if you have any reference files, it will get uploaded here and will be shown to you. Okay? So So we should inspect the skill before using it. We want to understand what it does and what input it expects and whether it's output is consistent with our requirements or not. So that we need to do all the time. Skills can help package repeatable task with But they don't automatically provide API access or allow an agent to modify external systems. If a scenario needs work automation connectors or external actions. We do have another platform which is called Microsoft Copilot Studio. That is the next option to evaluate. For today's session, the main takeaway is to start with the clear problem and only introduce extra capabilities when they add a real value. Microsoft documentation distinguishes Agent Builder's quick declarative agent experience from the more advanced action and workflow capabilities available through the Copilot studio. So, we have reached uh to the end of this today's session. We have built our Microsoft learn assistant, configured its instructions and knowledge source, tested its responses, and discussed how skills can help with repeatable task. Now, I would love to hear from you what kind of agent would you like to build? You can post your questions in the chat. Okay. So, let's discuss your ideas and I'll come up with that session in the future. So, before we finish, let's recap what we have learned. First, uh we started with the clear purpose for our agent. Second, we have written the instructions and we have explained how it should behave. Uh then we have configured relevant knowledge source instead of expecting the agent to know every details. And then fourth, we have tested the realistic questions. We have checked the supporting information and refined the configuration. And finally, we have added the capabilities only when our use case needs them. And you don't need to build a com- complicated agent on your first attempt. Start small, test it, and improve it. So, thank you so much for joining me today for this Share Tech Agent live with me session. I hope this practical walk-through has given you a clear understanding on Microsoft Copilot agent builder and the process of creating your own AI agents. I would encourage you to ex- experiment with a use case relevant to your work and explore the official Microsoft Learn documentation as you continue learning. Uh if you found this session useful, stay connected with SharePint for more practical sessions on Microsoft 365, SharePoint, Azure, and AI. Thank you for your participation and your questions. Keep learning, keep building, and keep sharing. Have a wonderful day, uh everyone, and yeah, thank you.

Assistir ao vídeo original no YouTube

Assuntos: Agentes de IA, Assistente pessoal, Assistente pessoal. Publicado no YouTube em 2026-10-18.

Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.

About Microsoft Copilot

Microsoft Copilot is an AI chatbot and virtual assistant published by Microsoft and developed by Microsoft with OpenAI. It was released on February 7, 2023 and is built on large language models.

Copilot is based on OpenAI technologies such as ChatGPT, GPT-4 and the DALL·E 3 image generator. It runs in web browsers, in Microsoft Edge and in Windows, and is offered under a proprietary license.

Fact sheet

• Product: Microsoft Copilot
• Type: Chatbot, virtual assistant, conversational AI
• Developers: Microsoft and OpenAI
• Publisher: Microsoft
• Release: 02/07/2023
• Based on: ChatGPT, GPT-4, DALL·E 3
• Platforms: Web browser, Microsoft Edge, Windows
• Official website: copilot.com

Read also

• Build Your Own AI Agent in Microsoft Copilot (No Coding Required!)
• ChatGPT Work: 10 AI Agents to Automate Your Work Right Now
• I Gave AI Agents Too Much

---

Sources and references: Wikidata (Q116793893) · official Copilot website · video by ShareTech on YouTube.