10.2 Creating a Copilot Studio Agent from a Canvas App

Agent builder: from app to agent

Agent builder in Power Apps creates a Microsoft Copilot Studio agent from an existing canvas app. You point it at the app and describe the business process you want to automate; agent builder then uses the app's metadata — its screens, data sources, and logic — together with your stated goal to generate step-by-step instructions, extract knowledge, identify triggers, and convert app capabilities into actions. The result is a fully equipped agent that you edit, test, and publish in Copilot Studio. This is the AB-410 blueprint skill 'create a Copilot Studio agent from a canvas app': creation starts inside Power Apps at make.powerapps.com, and refinement happens in Copilot Studio.

Prerequisites

• A tenant administrator must turn on Publish Copilots with AI features in the Power Platform admin center.
• The environment must include a Microsoft Dataverse database.
• Block unmanaged customizations must be disabled on the environment.
• The environment must have the Microsoft Copilot Studio solution, version 2.0.0.649.1 or higher.
• Agent builder is available only in regions where Copilot Studio generative agents are available, and the feature works only for canvas apps.

Entry points

1. On make.powerapps.com, select Agents in the left navigation pane (look under More if it is hidden), then select Create an agent from an app, pick the app, and select Next.
2. Alternatively, select Apps, choose your app, and select Create agent from app on the command bar — the same command appears under the app's Commands (...) menu.

A realistic scenario: the operations team has a canvas app for expense approvals, and the same three steps — read pending items, apply the threshold rules, update the status — consume an hour of someone's day. Instead of rebuilding that logic as a flow, the maker selects the app in agent builder and describes the outcome in plain language.

The creation flow

After you pick the app, agent builder offers suggestions for processes that could be automated, or you describe the goal yourself in the text box — for example, 'Review pending vacation requests and update their status based on predefined approval criteria.' Selecting a suggestion fills the text box automatically, and you can edit it before continuing. Agent builder then generates the step-by-step instructions plus the extracted knowledge and triggers. Review everything — you can edit any of it, or revise the description and select Regenerate instructions to get an updated set that matches the new goal. When satisfied, select Create and the agent is provisioned in Copilot Studio.

Writing instructions that work

Microsoft's guidance, and a likely exam theme:

• Use clear, specific wording — 'Filter the claims data using the approval status,' not 'Filter the data.'
• Keep each sentence simple; a single sentence overloaded with conditions may not execute as intended.
• Maintain logical flow — insert new steps in line in the correct position instead of appending them at the end of the instruction set.

Limitations to memorize

• If a data source connection is not found, actions cannot be extracted for the app — you get an agent with only the actions that could be converted, or none at all.
• Deprecated Excel connectors are unsupported; Excel Online (Business) and Excel Online are supported.
• Canvas apps only, and only in regions with Copilot Studio generative agents.

Edit, test, publish, share

Creation is not deployment. In Copilot Studio you refine the agent's instructions, knowledge, topics, and actions, validate behavior in the test pane, and then publish the agent so it can be shared with its intended audience. Sharing follows the agent's own security configuration: the agent acts through the connections and permissions it was given, and it does not grant users access to data they could not already reach. Because the agent is grounded in the app's screens, data, and logic, it can answer questions about the app's domain and perform the actions that were extracted from the app — end users invoke it through the channel where it is published rather than by opening Copilot Studio themselves.

The Copilot control: awareness level, plus a 2026 trap

Separately from agent builder, canvas apps have offered a Copilot control (preview) that lets app users chat with their Microsoft Dataverse data inside the player:

• The control supports only Dataverse tables as its data source.
• A Power Platform administrator must enable Allow users to analyze data using an AI-powered chat experience in canvas and model-driven apps in the Power Platform admin center.
• Makers turn on Copilot component, and optionally Edit in Copilot Studio for customization, under Settings > Updates > Preview.
• The control does not render in Studio preview — you must save, publish, and play the app in the web player to see it.

The trap: starting February 2, 2026, you cannot add the Copilot control to new canvas apps. Existing apps that already use it remain functional for a limited time, but Microsoft positions Microsoft 365 Copilot in canvas apps as the recommended replacement for natural-language interaction. If an exam option proposes adding the Copilot control to a brand-new app, treat it as outdated.

Governance and security

• The agent inherits the security posture of what it touches: app sharing determines who can use the app, and data permissions — Dataverse security roles and connector credentials — still gate the underlying rows. An agent grounded in an app does not bypass those controls.
• Administrators control the capability at the tenant level through the Publish Copilots with AI features setting, and environment-level settings such as block unmanaged customizations can silently prevent agent creation.
• Several pieces of this story are or were preview features: preview functionality is not meant for production use and carries supplemental terms, so know which features are preview before recommending them.

Quick contrast for the exam