Copilot Studio: Autonomous Agents

Copilot Studio has been the low code tool of choice for building conversational AI experiences that work with you, acting as your personal assistant to get answers and perform tasks.

As announced at Ignite this week, Copilot Studio has added a new set of preview capabilities that enable you to build autonomous agents that work for you.? These autonomous agents act as microservices that take desired actions on your behalf and respond to events without the need for human interaction.

You might be thinking: this sounds a lot like automation…

Well, you're correct!? Autonomous agents leverage automation to perform tasks.? However, these agents are foundationally different than typical automation.

The backbone of autonomous agents is the orchestrator, which uses generative AI to choose the best actions, knowledge, and topics to answer user queries or respond to event triggers.

Let’s say I want to create an agent that handles hardware procurement for new employees.

First, I need to add a trigger to my agent.? Unlike topic triggers, which require input from a user, event triggers allow your agent to act autonomously in response to the defined event occurring.

Whenever a new employee is created in our system, I receive an email requesting a device to be procured for the new employee.? I will add the Outlook “When an email arrives” trigger to my agent, so that it is triggered every time a new request comes in.

My next step is to add all the actions/workflows, knowledge sources, and topics that my agent needs to call.?

I will add my inventory spreadsheet as a knowledge source, so that the agent can pull the device options as well as their respective inventory counts.? I’ll also add the Microsoft Devices public website so the agent can get the device specs.

I will also add an action that runs my Power Automate workflow to submit the selected device into our procurement tool and a connector action for Outlook that sends a confirmation email to the new employee.

Now that I’ve given my agent all the pieces it’ll need to successfully run my process, I need to tell it how to use each piece.? This is where instructions come into play.

Agent instructions direct your agent on how to handle the information received from the trigger (new email request), what steps to take when triggered, and how to act.

In the instructions, I’m guiding my agent, telling it when to call different actions, when to get information from the knowledge sources, and when to reach out to a human for input.

I can now test my agent and follow the activity map to see how the process unfolds.

Then, I can click into each step and see what happened.

Note that I have not "hard-coded" the inputs for To, Subject, or Body. Instead, I used natural language in the instructions to tell the agent how to fill out these inputs. For example, my instructions tell the agent to send the email to the new employee - therefore, the agent is intelligent enough to fill in the "To" input for me.

As my agent processes these triggers, I can view the Activity tab to see each run.

Now that we've explored how it works, we can revisit the question, "how does this differ from traditional automation?"

Some of the benefits of using these autonomous agents features include:

Streamlined development and adaptability: Processes can be easily built and modified using natural language and the "building blocks" (actions, knowledge sources, and topics), simplifying implementation and updates

If my device inventory manager asks to be alerted whenever the inventory count for a device is less than 5, I can add a "Send an email" action and then modify my instructions to include that step.

Intelligent user interaction: Agents can seamlessly prompt users for input to address gaps in information, ensuring smoother workflows

I can instruct my agent to reach out to the new employee to make their device selection.

Enhanced flexibility for diverse use cases: These are nondeterministic agents that efficiently adapt to variations, making it easier to handle new and complex scenarios

A nondeterministic step could be added to my scenario that automatically selects the device with the higher inventory and lower demand when no device preference is specified in the email.

This hardware procurement agent is just one simple example of using these autonomous capabilities.

As Microsoft continues to invest in Copilot Studio and its autonomous capabilities, we will see a new wave of autonomous processing that relies less on the structure you give it and more on AI that intuitively understands your intent, adapts to changing needs, and delivers results with minimal human input.

Check out some of the additional Copilot Studio enhancements announced at Ignite!?

?

?