Copilot Agents: Build Your First Agent in Microsoft 365 Copilot

A Copilot Agent is a custom AI assistant that lives inside Microsoft 365 Copilot. Here's a best practice guide to building your first agent – from instructions to knowledge sources.

What is a Copilot Agent?

In short: a Copilot Agent is a custom AI assistant that lives inside Microsoft 365 Copilot.

Think of it like a "Project" in ChatGPT, a "Project" in Claude, or a "Gem" in Gemini. You give the agent a set of instructions, feed it relevant knowledge, and it responds within the domain you've defined. Every time. Without you having to repeat the context.

The difference from other platforms: Copilot Agents live directly in the ecosystem most organizations already use. Outlook, Teams, SharePoint, Word. IT has approved it. Security is in place. And your colleagues can use the agent without creating new accounts.

That's where the value lies: you build once, and the whole team benefits. Note: this feature typically requires a Copilot license, which is NOT cheap. Talk to your IT department about the "business case" for you and your team before spending too much on licenses!

Before You Start

You need a Microsoft 365 Copilot license. If your organization has Copilot, you automatically have access to Agent Builder.

How to find Agent Builder:

Open the Microsoft 365 Copilot app (copilot.microsoft.com).
On the left side, under "Agents", click New Agent.
You land in Agent Builder. Ready.

There are two ways to build: you can describe your agent in natural language (Describe tab), or configure manually (Configure tab). I recommend starting with natural language and then fine-tuning in the Configure tab.

Three Steps to Your First Agent

Step 1: Name and Description

Give the agent a short, precise name. Think of it as a tool your colleagues should be able to recognize if you're sharing it. "Onboarding Guide" is better than "My sweet onboarding agent v2".

The description tells Copilot (and your colleagues) what the agent is for. Keep it to 1-2 sentences. Example:

"This agent helps new employees find answers to questions about onboarding, practical procedures, and important internal guidelines."

Step 2: Instructions (System Prompt)

This is the most important step. The instructions are the agent's "system prompt". This is where you define:

What the agent should do
What it must not do
What tone it should use
Who the target audience is

You have up to 8,000 characters for instructions. Use them. The more precise you are, the better answers you get.

Example of good instructions:

You are an onboarding assistant for [Company Name]. Your primary task is to help new employees find answers to practical questions about their first time at the company. Always respond in a friendly and professional manner. Keep answers short and concrete. Always refer to specific documents when possible. If you don't know the answer, say so honestly and refer to the HR department. You must NOT give legal advice or salary negotiation advice.

The Trick: Meta-Prompting

Find it hard to write good instructions? Let the agent do it for you.

Use the Describe tab. Write in natural language what you want: "I want an agent that helps new employees with onboarding. It should answer questions about lunch arrangements, IT equipment, office hours, and internal policies."

Agent Builder automatically generates name, description, and instructions based on your input. Then go to the Configure tab and fine-tune. That's meta-prompting: you prompt the AI to write the prompt for you.

It works surprisingly well. And it saves you from the blank-screen feeling many get when writing instructions from scratch.

Step 3: Add Knowledge (Context Engineering)

This is where most people stop too early. An agent without knowledge is just Copilot with a title.

Context engineering is about giving the agent the right documents, data, and sources so it can actually respond accurately to the questions it receives.

You can add the following knowledge sources:

SharePoint files, folders, and sites (up to 100 files per agent, but stay well under that limit!)
Uploaded files directly from your computer
Public websites (up to 4 URLs, max 2 levels deep)
Teams conversations (channels, meetings, group chats, up to 5 links)
Emails from Outlook (requires Copilot add-on license)
Microsoft 365 Copilot Connectors (if your admin has enabled them)

Important detail: the agent always respects existing access permissions. If a colleague doesn't have access to a SharePoint document, they won't get the content through the agent either. It's a security feature, and it works — usually. Test thoroughly and report to Microsoft if the agent behaves strangely!

Five Agents You Can Build Tomorrow

1. The Onboarding Agent

Problem: New employees ask the same questions over and over. HR spends hours answering.

Solution: An agent that knows your employee handbook, IT procedures, lunch arrangements, and office policies.

Knowledge to add: Employee handbook (SharePoint), IT setup guide, welcome presentation, FAQ document.

Tip: Also add your organization chart. Then the agent can answer "Who is my nearest manager?" and "Who do I contact about payslips?"

2. The Project Status Agent

Problem: You spend 30 minutes gathering updates from Teams messages, emails, and documents before every status meeting.

Solution: An agent grounded in your project channel, relevant SharePoint files, and meeting notes.

Knowledge to add: Teams project channel, project plan (SharePoint), latest status reports, meeting minutes.

Tip: Be specific about which Teams channels you add. "All my chats" creates noise. A single project channel gives precision.

3. The HR Policies Agent

Problem: Employees ask HR about vacation, parental leave, sick leave, and benefits. The answers are in documents nobody reads.

Solution: An agent that has your personnel policies as its knowledge base and can answer specific questions.

Knowledge to add: Personnel handbook, vacation policy, parental leave rules, health insurance agreement, travel policy.

Tip: Set in the instructions that the agent must NOT give legal advice but should refer to HR when in doubt.

4. The Meeting Preparation Agent

Problem: You walk into meetings without having read the relevant documents. Everyone does.

Solution: An agent that can brief you on a specific topic or project in 30 seconds, based on the latest files and conversations.

Knowledge to add: Relevant SharePoint folders, Teams meeting history, project documentation.

Tip: Ask the agent to structure its response in three parts: (1) Key metrics, (2) Recent decisions, (3) Open questions.

5. The Product FAQ Agent

Problem: Sales and support staff spend time searching for product information across documents.

Solution: An agent that knows your product catalog, price lists, specifications, and competitor analyses.

Knowledge to add: Product datasheets, price lists, competitor comparisons, FAQ from support.

Tip: Add your company's website as a public source. Then the agent can also reference the official product page.

Best Practices: 8 Things That Make Your Agent Better

Be specific in instructions. "Be helpful" is too vague. "Answer in max 3 paragraphs, use a professional but informal tone, and always reference the document source" is precise.
Define what the agent must NOT do. Boundaries are as important as capabilities. "You must not give advice on salary, legal matters, or personal issues."
Test before sharing. Use the Try it tab in Agent Builder. Ask the questions your colleagues would ask. Even the weird ones.
Quality over quantity in knowledge sources. 100 irrelevant documents make the agent worse. 5-10 well-chosen and updated files make it good and precise.
Keep knowledge sources updated. The agent only knows what you give it. If your personnel handbook is from 2019, you get answers from 2019.
Use starter prompts. Add 3-5 suggested questions so users know what the agent can do.
Share with the right people. Recipients must also have access to the underlying SharePoint files.
Iterate. Your first version won't be perfect. Use it for a week, note what's missing, and update.

The Four Most Common Mistakes

No knowledge sources. An agent without context is just a worse version of Copilot. Add documents.
Too vague instructions. "Be helpful" is meaningless for an AI. Be concrete: what should it do, for whom, in what tone, with what limitations.
Too many knowledge sources. If you dump your entire SharePoint drive in, the agent drowns in data. Be selective.
Never updated. An agent with outdated files from last year gives outdated answers. Set a reminder to review quarterly.

Agent Builder vs. Copilot Studio

In short: Agent Builder is for you and your team. Copilot Studio is for the IT department.

Use Agent Builder when:

You quickly want to build an agent for yourself or a small team
The agent should use existing M365 data (SharePoint, emails, Teams)
You don't need advanced workflows or external integrations

Use Copilot Studio when:

The agent should be used by an entire department or organization
You need multi-step workflows, API integrations, or custom connectors
IT requires advanced governance and deployment control

The good news: you can start in Agent Builder and later copy your agent to Copilot Studio if you need more.

Human in the Loop: You're the One in Control

No matter how good your agent gets, there's one thing you must never forget: you are the one who's accountable.

A Copilot Agent is a tool – not a decision-maker. It can summarize, suggest, and execute tasks within the boundaries you've defined. But it doesn't understand context the way you do. It doesn't know the political dynamics in your team. It doesn't know that one colleague has had a rough week. It can't judge whether an answer is good enough to send to a client.

That's why human in the loop isn't just a buzzword – it's a design principle:

Review the agent's output before you act on it. Especially for client-facing responses, legal questions, or decisions with consequences.
Set boundaries in your instructions. Explicitly define what the agent must not do. "You must not give definitive answers on legal questions" is a safety valve.
Use the agent as a sparring partner, not an oracle. The best use of AI is when it gives you a better starting point – not when it replaces your judgment.
Build feedback loops. When the agent answers incorrectly, note it. Update instructions and knowledge sources. That's how the agent improves over time – through your guidance.

Remember: AI agents amplify what you give them. Good instructions, good data, and human oversight produce good results. Poor guardrails produce poor results – faster than ever before.

We humans control the machines. Not the other way around. And that's a conscious choice you need to make every time you build and use an agent.

Quick Checklist: Before You Publish Your Agent

Does the agent have a clear, recognizable name?
Is the description short and precise (1-2 sentences)?
Are the instructions specific? (What should it do / not do / tone / audience)
Have you added relevant knowledge sources (SharePoint, files, websites)?
Have you tested with realistic questions in the Try it tab?
Have you tested with questions OUTSIDE the agent's scope?
Have you added starter prompts so users know what they can ask?
Do recipients have access to the SharePoint files the agent uses?
Have you set a date for the next review of the agent?

The most important thing to remember: You don't need to be a prompt expert to build a good agent. You need a clear idea of what it should solve, the right documents as context, and 15 minutes.

Start small. Build one agent. Use it for a week. Learn from it. Build the next one.

🎓 Want hands-on training? Book our Copilot Agent Workshop — from idea to deployed agent in half a day.

📬 Subscribe to the newsletter "AI, Built Human" on Substack — weekly insights on AI in practice.

Stefano Vincenti · AI Advisor & Trainer · aitrainer.dk · External Lecturer, IT University of Copenhagen · Cofounder & CTO BotTellMe · Partner, TryZone