--- id: copilot-agent-builder lang: en track: copilot title: "Agent Builder" seoTitle: "Agent Builder: build a Copilot agent" description: "Use Agent Builder to turn a repeatable task into a no-code Copilot agent with its own instructions and knowledge sources, then share it safely with your team." part: working-smart partTitle: "Working smart" partOrder: 4 order: 2 level: Intermediate durationMinutes: 7 outcome: "Turn a repeatable task into a Copilot agent you open on purpose and can share with your team." summary: "Agent Builder packages a repeatable task into a standalone, no-code assistant you open on purpose, unlike a Claude Skill that fires quietly mid-chat, and you can share it with your team." draft: false updated: "2026-10-14" furtherReading: - title: "Build agents in Agent Builder" url: "https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/agent-builder-build-agents" note: "Microsoft's step-by-step guide to creating an agent." - title: "Add knowledge sources to your declarative agent" url: "https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/knowledge-sources" note: "Which knowledge sources an agent can draw on." - title: "Share and manage agents built in Agent Builder" url: "https://learn.microsoft.com/en-us/microsoft-365/copilot/extensibility/agent-builder-share-manage-agents" note: "Sharing roles, the Agent Store, and managing the agents you built." - title: "Build your own agent with Microsoft Copilot" url: "https://support.microsoft.com/en-us/microsoft-365-copilot/build-your-own-agent-with-microsoft-365-copilot" note: "Microsoft's end-user introduction to building your own agent." keywords: - Copilot Agent Builder - build a Copilot agent - no-code agent - agent knowledge sources - share a Copilot agent - agent instructions quiz: intro: "Five quick questions. Ace them all and the next lesson unlocks." passScore: 5 questions: - q: "What does Agent Builder actually create?" options: - "A script that runs on a schedule" - "A declarative agent: a custom version of Copilot with your instructions and chosen knowledge sources, built without code" - "A new response mode inside standard Copilot Chat" - "A copy of Copilot that can read every file in your tenant" answer: 1 explain: "A declarative agent: a custom version of Copilot that follows your written instructions and draws on knowledge sources you chose, built without code." - q: "How is using a built agent different from how a Claude Skill fires?" options: - "There is no difference: same feature, different name" - "A built agent steps in automatically mid-chat, while a Claude Skill has to be opened" - "A built agent only works inside Outlook" - "A Claude Skill is matched automatically from your wording; a built agent is a standalone destination you open on purpose" answer: 3 explain: "A Claude Skill is matched automatically by Claude based on your wording; an Agent Builder agent is its own standalone destination you or others open on purpose, not something that quietly steps into your regular chat." - q: "Which field does the most to help Copilot recognise when your agent fits a task, and also doubles as what people see before opening it?" options: - "The description" - "The name" - "The starter prompts" - "The knowledge sources" answer: 0 explain: "The description. Keep it short, precise, and specific rather than generic." - q: "A colleague has \"Can chat\" access to an agent you built. What can they do?" options: - "Edit its instructions and knowledge sources" - "Add or remove other people and change sharing" - "Chat with the agent, but not change its instructions, knowledge sources, or sharing settings" - "Nothing, until an admin approves them" answer: 2 explain: "Use the agent to chat with it, but not edit its instructions, knowledge sources, or sharing settings. Only \"Can edit\" owners get that." - q: "You share an agent that reads from a specific SharePoint folder. Does everyone you share it with now see that whole SharePoint site?" options: - "Yes, sharing an agent shares everything it can see" - "No. Only the files and sites you added as knowledge sources travel with the share, and each user's own permissions still apply" - "Yes, but only for people with Can edit rights" - "Only if the agent is listed in the Agent Store" answer: 1 explain: "No. Only the files and sites you explicitly added as knowledge sources travel with the share, and the agent still respects each individual user's own permissions." pass: badge: "Perfect score" title: "You can package a job into an agent." body: "Full marks. You know what Agent Builder makes, how it differs from a Claude Skill, and how to share it safely. Next up: Copilot Notebooks, a scoped home for work you'll keep coming back to." retry: "Check the marked answers and give it another go. You've got this." --- > **Level:** Intermediate ยท **Reading time:** ~7 min > > **What you'll be able to do:** turn a repeatable task into a Copilot agent you open on purpose and can share with your team. Notice how often you type the same setup into Copilot? "Summarise this against our QBR template, flag anything missing, keep it under a page." Every time, from the top. Agent Builder is how you stop retyping it: build the job once, and hand it a name.
Matched versus opened A Claude Skill is matched automatically by Claude based on your wording, with no selecting needed. An Agent Builder agent is its own standalone destination that you or a colleague opens on purpose, the way you would open any other chat. Not matched. Opened. Claude Skill You describe your task in your own words. Claude matches it and opens the Skill quietly. You never select it. Agent Builder agent Its own destination, under My agents or the Agent Store. You open it on purpose. Shareable with a team.
## What Agent Builder actually makes Agent Builder creates what Microsoft calls a declarative agent: a custom version of Copilot that follows instructions you wrote and draws on knowledge sources you chose, without anyone writing a line of code. You reach it from the Agents section of Microsoft 365 Copilot, and you can build one three ways: describe what you want in plain language and let Copilot draft the configuration, fill in the fields yourself on the Configure tab, or start from a ready-made template. ## Not matched, opened Here's the real difference from a Claude Skill, and it's worth getting straight rather than assuming the two work the same way with different names. A Claude Skill sits quietly on a shelf. You never select it: you just describe your task in your normal words, and Claude matches it to the right Skill's description and opens it behind the scenes. You would barely know it fired. An agent you build in Agent Builder doesn't work like that. It's not something Copilot reaches for mid-conversation while you're typing in your regular chat. It's its own standalone destination, listed under My agents or found in the Agent Store, that you or whoever it's shared with opens on purpose, the way you'd open any other chat. Ask a question in ordinary Copilot Chat and your custom agent won't quietly step in to answer it: you have to go to the agent itself. That makes an agent closer to a dedicated assistant you built than a reflex Copilot develops. Picture the difference side by side. With a Claude Skill, you type "summarise this against our QBR template" into your normal chat and Claude quietly recognises the match and applies it; you'd struggle to say exactly when it kicked in. With an Agent Builder agent, you navigate to "QBR Summariser" in your Agents list first, then ask your question there. One is a reflex baked into the conversation you're already having. The other is a door you walk through on purpose. Neither is wrong, they're just different shapes for the same underlying idea: package a task so you don't re-explain it. ## Building one Four fields do the real work. The name is short and unique, thirty characters at most, so pick something that says what it does at a glance. The description matters more than its length suggests: it's what Copilot uses to recognise when the agent fits a task, and it's also the blurb people see before they open it, so keep it short, precise, and specific rather than generic. The instructions are the real brief, up to eight thousand characters of plain-language direction on the task and exactly how to do it; describe the agent in natural language and Copilot writes a first draft of these for you. Knowledge sources are what the agent is allowed to read from: specific files, SharePoint sites, Teams chats, or public web pages, depending on what your licence allows. A fifth field, starter prompts, isn't strictly required but earns its keep: a few example questions with short descriptions that show anyone opening the agent what it's actually good for, so they're not staring at a blank box guessing what to type. Before you finish, the Try it tab lets you test the agent against a real question and see what comes back. If the first draft needs a tweak, you don't have to start over: describe the change in plain language, something like "update it to summarise Teams chats instead of emails," and the configuration updates to match. ## Sharing is where this pulls ahead of a Skill A Claude Skill you build stays with your own account. If three people on your team all want it, all three build or upload their own copy. An Agent Builder agent doesn't have that limit, and it's a genuine strength worth using on purpose rather than an afterthought. Every new agent starts private, visible only to you. From there you can add people by name, group, or email, and give each one a role: Can edit, which makes them a full owner who can add or remove other people and adjust sharing, or Can chat, which lets them use the agent without being able to change it. There's also an org-wide toggle that lists the agent in your tenant's Agent Store for anyone to start chatting with, no individual invitation required. If your organisation restricts who can turn that on, an admin reviews and deploys the agent before it goes any wider, and an agent aiming for the most visible spot, the Built by your org section of the store, goes through that same admin review. One safeguard carries straight over from grounding in [lesson 4.1](/learn/copilot/working-smart/context-and-grounding): sharing an agent never overrides anyone's own permissions. Only the files and sites you explicitly added as knowledge sources travel with the share, not blanket access to your SharePoint, and whatever the agent later reads from a connector still respects the individual permissions of whoever is chatting with it. Sharing the agent is not the same as sharing everything it can see. ## Where it earns its keep Build an agent for a task that's repeatable and has a clear right way to do it: a document-review checklist your whole team should run the same way, a report format with a fixed shape, a Q&A assistant grounded on one specific set of policy documents. These are jobs where consistency is the entire point, and where handing the finished agent to three colleagues beats explaining the same instructions three separate times. Skip it for a one-off question or open-ended exploration. If you'll never ask the same kind of thing again, a plain prompt in ordinary Copilot Chat is faster to write and doesn't need a name, a description, or anyone's permission to open. A useful test: if you've typed a near-identical prompt three times this month, or you're about to explain the same instructions to a second colleague, that's usually the moment an agent pays for itself. Before that point, it's just overhead. ## One safety note An agent you share carries real weight: its instructions shape how it answers, and its knowledge sources shape what it can see. Treat building one for other people the way you'd treat handing them a document with your name on it: read your own instructions back before you share, keep the knowledge sources narrow and intentional rather than adding "everything, just in case," and remember that giving someone Can edit rights hands them equal control over the agent, not a lesser copy of it. ## Where this goes next An agent is the *what*, a job done your way, available to open on demand. Next, Copilot Notebooks gives ongoing work a scoped home that answers only from the sources you chose. The two lessons after that cover the *who else helps*: Researcher and Analyst, Copilot's own prebuilt agents for research and data work, and Connectors, which widen what any agent, built-in or your own, can actually see. ## Try it Pick one task you keep re-explaining to Copilot, the kind with a clear right way to do it. Sketch its agent in four lines: a name, a one-line description sharp enough to tell a colleague what it's for, three or four instructions in order, and one knowledge source it should actually need. You don't have to build it yet. Seeing a real task from your week break cleanly into those four pieces is the point, that's the moment an agent stops being abstract. ## Key terms [Agent Builder](/glossary#agent-builder), [Agent](/glossary#agent), [Grounding](/glossary#grounding), [Knowledge source](/glossary#knowledge-source), [Microsoft Graph](/glossary#microsoft-graph), [Copilot](/glossary#copilot)