Build agents by using Agent Builder in Microsoft 365 Copilot

The Agent Builder feature in Microsoft 365 Copilot provides a simple interface that you can use to quickly build declarative agents by using natural language. This article describes how to use Agent Builder to build an agent for Copilot.

When you choose New agent in Microsoft 365 Copilot, you can:

• Use natural language to describe your agent, and Copilot builds it for you (recommended).
• Configure your agent manually via the Configure tab.
• Start with a template to create an agent for a specific use case.

Use natural language to describe your agent (recommended)

You can use Agent Builder in Microsoft 365 Copilot to create an agent by using natural language. As you provide information conversationally, the agent’s name, description, and instructions update automatically to refine its behavior. This approach:

• Understands a user's intent through natural language.
• Provides suggestions, guidance, and next-best prompts.
• Generates optimized agent instructions to build accurate, high-quality agents.

Using natural language saves time and helps you create effective agents without manual configuration. In response to your natural language description, Agent Builder:

• Configures the agent automatically.

• Suggests and applies refinements. For example: "Include email attachments in summaries". Agent Builder suggests:

  • Show only attachments from unread messages
  • Summarize attachments by file type

• Adds knowledge sources. For example: "Summarize my daily emails and include attachments." Agent Builder adds My emails as a knowledge source.

• Integrates capabilities. For example: "Create a PowerPoint presentation from this outline" or "Analyze this Excel data and generate a chart." Agent Builder configures your agent with code interpreter and image generator.

You can also choose the plus sign (+) in the chat box to select and add or search for knowledge sources.

When you finish describing your agent and adding knowledge sources, view the configuration on the Configure tab. The tool prepopulates:

• Name
• Description
• Instructions (including guidelines, steps, error handling, examples)
• Knowledge sources
• Suggested prompts

After you create your agent, test it on the Try it tab. You can continue to refine it by using natural language. If you want to change the starter prompts or intent, just say it in natural language - for example: "Update the agent to summarize Teams chats instead of emails."

Use the Configure tab to create your agent manually

If you don't want to use natural language and let Agent Builder configure your agent for you, you can configure your agent's behavior manually on the Configure tab.

To configure your agent manually:

1. In Microsoft 365 Copilot, select New agent.
2. On New agent, select Skip to configure.
3. Agent Builder opens the Configure tab, where you can specify your agent's name, description, instructions, knowledge, and prompts.

The following table describes the fields that make up the agent.

After you create your agent, test it on the Try it tab. You can continue to refine your agent's instructions, knowledge, and starter prompts.

Build from a template

Agent Builder in Microsoft 365 Copilot includes templates that you can use to build agents for specific use cases. The templates come preconfigured with a description, instructions, and prompts. Use the templates as-is or customize them for your specific needs, such as adding more knowledge sources and capabilities.

To use a template to build your agent:

1. In Microsoft 365 Copilot, select New agent.
2. On the New agent page, under Start with a template, choose the template you want to use. The agent is created automatically.
3. You can further refine your agent and add knowledge sources by using natural language, or add your refinements manually on the Configure tab.

For more information, see Agent templates.

After you create your agent, test it on the Try it tab. You can continue to refine your agent's instructions, knowledge, and starter prompts.

Add knowledge sources

To build context-aware agents, reference SharePoint items or any public websites. If your users have a Microsoft 365 Copilot add-on license, you can also:

• Ground your agents in personal work information, such as Teams chat messages and Outlook emails.
• Use prebuilt Microsoft 365 Copilot connectors that are enabled in your tenant.

Choose knowledge sources to add to your agent in the chat box on the Describe tab if you're using natural language to create your agent, or select knowledge sources on the Configure tab.

For more information, see Add knowledge sources.

Add capabilities

Use natural language to add capabilities on the Describe tab, or configure them in the Capabilities section of the Configure tab. You can add the following capabilities to your agent:

• Code interpreter - Solves complex math problems, analyzes data, and generates visualizations. To add this capability, select the toggle next to Create documents, charts, and code.
• Image generator - Generates images based on user prompts. To add this capability, select the toggle next to Create images.

Test your agent

The Try it tab provides an instance of the agent that you can use to test and refine your agent within the authoring experience. Use it to test the agent while you're creating or making updates to it. The Try it experience is enabled after the agent has a name, description, and instructions. It behaves according to those instructions like a published agent, including the ability to respond to complex queries within its realm of specified knowledge. The agent updates during each turn of the conversation as you add new information by using natural language or manually on the Configure tab in Agent Builder.

Some agent features aren't available in the Try it experience. For example, you can't use it to share prompts, provide feedback, or @mention other agents in the Microsoft 365 Copilot app. You can use these features after you create the agent.

The Try it experience includes suggested starter prompts, which when selected invoke the prompt and start the conversation. You can select New chat to start a new conversation with the agent to view the starter prompts again.

Next step