Build Your First Autonomous Agent with Copilot Studio

How to Build Your First Autonomous Agent

Introduction

• Overview of building an autonomous agent using Co-Pilot Studio.

• Focus on use cases and reasons for creating autonomous agents.

• Scenario involves a customer-facing returns policy for a retailer.

Building the Agent

• Aim to create an agent that processes return requests autonomously.

• Emphasis on moving beyond chat experiences to handle workflows independently.

• The project will not include human escalation; it will be fully autonomous.

Getting Started with Co-Pilot Studio

• Introduction to Co-Pilot Studio as a no-code tool for building agents.

• Free trial available; no license needed for testing and building.

• Starting by creating a new agent named "returns agent."

Configuring the Agent

• Importance of providing clear descriptions for the agent's purpose.

• Description should outline the agent's role in handling return requests.

• Avoid creating overly complex agents with multiple responsibilities.

Adding Knowledge and Instructions

• Each agent should focus on being an expert in a specific domain or process.

• Plan to upload knowledge files after provisioning the agent is complete.

How Does an Autonomous Agent Work?

Understanding the Process

• An autonomous agent automates processes without traditional if-then logic, using knowledge instead.

• It utilizes a returns policy and can send emails to customers based on their inquiries.

• The trigger for actions can be an incoming email or other forms like web forms or phone calls.

Enabling Generative AI Features

• Generative AI is activated to determine responses, moving away from traditional topic triggers.

• The agent uses descriptions to decide actions without a predefined flowchart, mimicking human decision-making.

• Ensure generative AI capabilities are enabled in settings for optimal functionality.

Adding Knowledge Sources

• Upload the returns policy document as knowledge; it must be closed before uploading.

• Uploaded documents are indexed for AI reasoning but require updates if changes occur externally.

• Other options include connecting to websites or SharePoint for internal use cases.

Managing Knowledge Scope

• Limit the AI's general knowledge usage by disabling its ability to answer questions outside the returns policy.

• This ensures that responses are strictly based on provided information rather than general assumptions.

Testing the Agent

Getting Started with Generative AI

• The speaker discusses activating generative AI capabilities and testing functionality step by step.

• Demonstrates returning an item using the correct returns policy from a single knowledge source.

• Introduces the concept of enabling actions for the agent to make decisions based on knowledge.

Implementing Actions

• Describes how to set up actions similar to Power Automate, focusing on sending emails.

• Explains that flows can be used for retrieving information but focuses on a simple email action.

• Mentions using Outlook connector for sending emails and the need for authentication.

Configuring Email Action

• The action is renamed to "send an email" for clarity in instructions provided to the agent.

• Emphasizes clear descriptions for actions, detailing what the email will communicate regarding returns.

• Discusses options for escalating emails to a human before sending them out.

Setting Up Email Parameters

• Inputs required include recipient, subject, and body of the email; dynamic filling is possible.

• Shows how to set specific values like subject line while allowing generative AI to handle content.

• Confirms minimal input requirements while leveraging generative AI capabilities effectively.

Testing Email Functionality

• Reviews setup steps confirming generative mode must be enabled for functionality.

• Prepares to test by providing specific instructions related to item return queries via email.

How to Automate Email Responses

Overview of Email Automation

• Discusses the initial setup for automating email responses using Flybys dollars.

• Emphasizes the ease of populating emails without needing a complex flow; encourages questions for better understanding.

Instructions and Triggers

• Introduces the concept of giving instructions in natural language instead of building a traditional flow.

• Explains how to instruct an assistant-like system to handle customer inquiries effectively.

Task Execution

• Details on pasting instructions that guide the agent through tasks similar to human interaction.

• Mentions changing subject lines for demo purposes and outlines tasks to perform upon receiving an email.

Finalizing Instructions

• Suggests experimenting with wording in instructions, providing a copy for reference.

• Highlights key components like friendly greetings and avoiding citations in email responses.

Adding Triggers

• Describes adding triggers to automate processes without manual input, familiarizing users with Power Platform connectors.

• Discusses setting parameters for incoming emails, such as inbox selection and subject filters.

Testing Automation

• Outlines steps taken during testing, including sending a test email with specific criteria.

Understanding Email Flow Automation

Overview of the Flow

• The flow is designed to trigger when a new email arrives, utilizing specific parameters like email address and subject filter.

• Editing the flow may cause issues; it's advised to observe without making changes initially.

• Refreshing the run history shows no runs yet; patience is required as it typically takes about 2 minutes for triggers.

Voice Interaction Capabilities

• The system can also function as an IVR or voice chat agent, allowing users to interact via phone calls.

• It utilizes voice recognition to understand queries and respond appropriately, enhancing user experience.

Testing the Trigger

• A mistake in the subject filter was identified; testing with a corrected demo email was initiated.

• Real-time testing is emphasized; refreshing will show updates in Power Automate slightly ahead of other interfaces.

• The trigger successfully processed an incoming email, demonstrating its functionality without manual input.

Finalizing and Publishing

• The process includes checking knowledge sources and confirming that responses are generated correctly based on received emails.

• Formatting improvements are still needed for better presentation of automated responses.

How to Publish a Chatbot?

• The chatbot is published without needing an external facing experience; it runs as an automated flow.

• The system autonomously waits for emails that meet specific criteria to process requests.

• A draft email requesting a return of an item purchased with Afterpay is prepared before sending.

Real-Time Email Processing

• The agent receives the email trigger and processes it, aiming for near real-time response.

• The email processing involves checking knowledge sources and making decisions based on the content.

• There are plans for more tutorials on autonomous agents, including travel policy checks.

Learning Journey with Autonomous Agents

• This tutorial showcases one action from one knowledge source, highlighting the potential for expansion.

• Feedback and ideas for future tutorials are welcomed as part of the learning journey.