Power Platform Cloud Solutions Architect @ Microsoft | Microsoft BizApps MVP 2023 | Power Platform | SharePoint | Teams
Copilot Agent Builder lets you create a no-code agent that answers questions using information from selected files, rather than relying on a broad, general-purpose chat. It is useful for Microsoft 365 Copilot users who need a focused assistant for a recurring set of questions. Damien Bird demonstrates the setup with an example whisky guide grounded in a SharePoint folder.
An agent combines a defined purpose, instructions and selected knowledge sources to answer questions in a particular context. In this example, the agent is called Glen Damo Whisky Guide and uses a linked SharePoint folder containing sample business files. The intended result is a conversational way to ask about information in those files, such as the price of a sherry cask.
The approach is different from asking a general Copilot question without specifying a knowledge base. The creator chooses the information the agent should use and can also set whether other sources are available. In the example, Outlook, Teams and web search are switched off, leaving the linked folder as the agent’s knowledge source.
That source selection matters because a useful answer depends on the agent having access to the relevant, up-to-date information. The example also shows that adding information to a file does not automatically guarantee that the agent will interpret it as intended: when new non-whisky gifts are added to a workbook, the agent initially misses them. An instruction clarifying the workbook’s contents addresses the issue.
This workflow is aimed at people who have a recurring set of questions and a collection of documents that can answer them. Departmental subject-matter experts, team leads and other Microsoft 365 users can use the same general pattern for a focused knowledge assistant, without writing code. The example focuses on a small team sharing answers about products and prices.
A Microsoft 365 Copilot license is listed as a requirement for the demonstrated setup. The example uses a SharePoint folder as the knowledge source and shares the resulting agent with a team. Before building one for work, identify the folder that holds the information and confirm that it is the right place to maintain the source content.
The description is also a practical prerequisite for testing in Preview: Preview will not test the agent without one. Start with a clear description of the agent’s purpose and audience, then check the configuration before relying on responses.
The setup follows a compact sequence: define the agent, connect its knowledge, constrain its behavior and test questions against real content. Agent Builder lets you make these choices without a coding environment. The steps below capture the configuration used for the example agent.
Testing should include questions that the source files clearly answer, as well as questions that expose gaps or ambiguous wording. The gift example is a useful reminder to test after source files change: an instruction can help explain how information is organized, but the agent still needs to be checked against the updated content.
The example is a file-grounded agent for answering questions, not an agent configured to take actions. Its scope depends on the connected information and the instructions that explain how to use it. If a file does not contain an answer, or if its contents are unclear to the agent, users should not assume the agent will infer the intended result correctly.
Keeping the source material current is part of operating this kind of assistant. When a workbook changes, test the questions that depend on it again; the example shows that a newly added product was initially missed. Clear instructions about the role of each file can help, but testing remains a practical check before relying on an answer.
Linking a folder is presented as the preferred approach in this example, rather than uploading files one by one. That makes the folder itself the focus of the connection, but it also makes it important to select the intended location and check the knowledge access shown when sharing. The example highlights this access check as a practical step.
The illustrated business, people and files are fictional sample data, so the specific whisky prices and product details are examples, not real business information. The useful takeaway is the build pattern: select a bounded knowledge source, explain its contents, turn off irrelevant sources, and test representative questions before sharing.
The demonstrated setup lists a Microsoft 365 Copilot license as a requirement.
Yes. The example links a SharePoint folder as the agent’s knowledge source and tests it with questions about the files.
Preview will not test the agent without a description.
Use the agent’s instructions to explain which file contains what information, then test questions against the workbook.