Create an agent using the new Copilot Studio experience

Scenario

In this exercise, you will:

• Switch to the new Copilot Studio experience
• Create an agent using natural language instructions
• Add a prebuilt action the agent can use autonomously
• Test the agent and observe its reasoning behavior
• Refine instructions based on test results

This exercise will take approximately 30 minutes to complete.

What you will learn

• How the new Copilot Studio experience differs from the classic interface
• How to configure an agent using instructions instead of topic trees
• How the agent reasons autonomously to select knowledge and actions
• How to iterate on instructions to improve agent behavior

High-level lab steps

• Switch to the new Copilot Studio experience
• Create an agent with a description
• Write and refine agent instructions
• Add a prebuilt action
• Test and iterate

Prerequisites

• Have a Microsoft Entra ID account
• Have a Copilot Studio license or have signed up for a free trial.
• Have access to a Power Platform environment and a solution where you can create agents and related assets.
• You can use:
  • the environment and Lab Exercises solution created in the ILT Setup lab, or
  • your own existing environment and solution.
• If you do not already have an environment and solution prepared, complete the steps in the ILT Setup lab before continuing.

[!IMPORTANT] The new Copilot Studio experience is the redesigned interface that replaces the classic topic authoring canvas. Steps and screenshots in this lab reflect the new experience. If your environment shows the classic interface, follow the prompt to switch to the new experience before proceeding.

Key concept: Instruction-driven agents

The new Copilot Studio experience replaces the topic authoring canvas with a simpler, instruction-driven model:

The new experience is best suited for agents that need to handle open-ended, context-dependent conversations. The classic experience (covered in later labs) remains the right choice when you need a guaranteed, auditable sequence of steps — for example, collecting structured data in a specific order.

Exercise 1 - Switch to the new Copilot Studio experience

Task 1.1 – Open the new experience

1. Navigate to the Copilot Studio home page at https://copilotstudio.microsoft.com/ and sign in if prompted.

2. At the top of the page, verify that you are working in the environment you want to use for this exercise.

3. If you see a banner or notification prompting you to Try the new experience, select it.

   Alternatively, if you do not see the New experience toggle, navigate directly to https://copilotstudio.preview.microsoft.com/ to open the new experience.

   [!NOTE] The new experience may be the default in your environment. If the interface already matches the steps in this lab, you are using the new experience and can proceed to Exercise 2.

Task 1.2 – Review the new interface

Take a moment to review how the new experience is organized before creating your agent.

1. In the left-hand navigation, note the main sections: Home, Agent Ops, Chat, Agents, and Workflows.

2. The Agents page shows any existing agents. Agents created in the new experience do not have a topic canvas — their behavior is defined entirely by instructions, knowledge, and actions on the agent’s Overview page.

3. Notice there is no Topics section in the navigation. In the new experience, topics are replaced by actions and instructions.

Exercise 2 - Create an agent

In this exercise, you will create an IT support agent for a fictional company called Contoso. The agent will help employees troubleshoot common IT issues and submit support tickets.

Task 2.1 – Create the agent

1. Select Home in the left-hand navigation.

2. On the Home page, enter the following description in the build prompt text box:

   You are an IT support agent for Contoso. You help employees troubleshoot common IT issues such as password resets, software installation problems, and network connectivity. When you cannot resolve an issue, you help the employee submit a support ticket.
   

   [!NOTE] The new experience uses your description to automatically generate a set of initial instructions. Review them before continuing.

3. If prompted with clarifying questions, answer them to help the agent generate more accurate instructions. For example:

   • If asked how employees submit IT support tickets, choose the option to send a support request email to the IT helpdesk (rather than ServiceNow, Jira, Dynamics 365, or Zendesk).
   • If asked whether the agent should reference a knowledge base, SharePoint site, or website, choose the option indicating no external knowledge source is needed.

   Select Continue to advance through the questions, or Skip any question that does not apply. Select Submit when finished.

   [!NOTE] Depending on your description and environment, the clarifying questions may vary or may not appear at all. Answer them in line with the lab scenario rather than expecting these exact questions or answer options.

4. If prompted for the email address the agent should use when sending support requests, enter your admin email address, then select Submit.

5. Review and accept the auto-generated instructions and create the agent.

6. In the Artifacts panel on the right, select your agent to open its Build tab.

Task 2.2 – Review and refine the auto-generated instructions

1. On the agent’s Build tab, locate the Instructions section.

2. Review the instructions that were generated from your description. They should describe the agent’s purpose, tone, and general behavior.

3. In the Instructions section, update the instructions by adding the following guidelines, then select Save:

   ## Guidelines
   - Always respond in a professional and friendly tone.
   - For password reset requests, direct the employee to the self-service portal at https://aka.ms/sspr before offering to raise a ticket.
   - For issues you cannot resolve, collect the employee's name, email address, and a brief description of the issue before submitting a ticket.
   - Do not speculate about hardware failures. Always recommend contacting the IT desk directly for physical hardware issues.
   - When an employee's issue cannot be resolved, use the Send an email action to notify the IT helpdesk at helpdesk@contoso.com with the employee's name, email, and issue description.
   

   [!NOTE] Instructions in the new experience are the primary way to control agent behavior. Well-written instructions reduce the need for additional configuration and make the agent more predictable.

Exercise 3 - Test and refine

In this exercise, you will test the agent and observe how it reasons before responding.

Task 3.1 – Open the Preview tab

1. Select the Preview tab at the top of the page to test the agent.

2. When you test the agent, a reasoning summary appears automatically above each response, describing how the agent decided what to do. Select Show more on that summary to expand the full reasoning trace.

   [!NOTE] The reasoning trace is a key feature of the new experience. It shows which knowledge sources or actions the agent considered and why — for example, a Loaded Skill entry indicates which action or skill the agent invoked.

Task 3.2 – Test the instructions

1. At the top of the Preview tab, select New chat.

2. Enter the following prompt:

   I forgot my password and cannot log in.
   

   Based on the instructions you wrote, the agent should direct you to the self-service portal before offering to raise a ticket.

Task 3.3 – Test the agent’s reasoning behavior

1. At the top of the Preview tab, select New chat.

2. Enter the following prompt:

   My laptop will not turn on at all.
   

   Based on the instructions you wrote, the agent should decline to troubleshoot the hardware failure remotely, recommend contacting the IT desk directly. The agent might also offer to submit a support ticket on your behalf.

3. If prompted, provide your name, email, and a brief description of the issue.

Task 3.4 – Refine instructions based on test results

1. Review how the agent responded across the three test sessions.

2. If any response was not aligned with the intended behavior, select the Build tab, and in the Instructions section, adjust the relevant guideline.

   For example, if the agent did not mention the self-service portal for password resets, make the instruction more explicit:

   - For ALL password-related requests, always mention https://aka.ms/sspr as the first step before any other assistance.
   

3. Select Publish, then select Save and Publish, and re-test the affected scenario.

   [!NOTE] Iterating on instructions is the primary tuning mechanism in the new experience. Small changes in wording can significantly change agent behavior.

Summary

In this lab, you used the new Copilot Studio experience to create an instruction-driven IT support agent. You configured behavior entirely through natural language instructions and added a prebuilt connector action directly in Copilot Studio — without building a Power Automate workflow or leaving the page. Compared to Lab 03, where adding a tool required authoring a workflow, configuring inputs and outputs, and publishing it separately, the new experience significantly reduces the authoring effort for straightforward actions. You also used the reasoning trace in the Preview pane to observe how the agent decided what to do before generating a response. Having worked with both the classic and new experiences, you can now choose the right approach for each scenario: instruction-driven for open-ended conversations, and classic topics and workflows when you need a guaranteed, auditable sequence of steps.