Introducing Copilot Agents

In last week’s Friday edition of the Innovation Profs newsletter, we introduced Copilot agents, which are custom versions of Copilot tailored to a specific task. We surveyed a number of prebuilt agents, including Researcher, Analyst, and Learning Coach.

This week, we’ll walk through the steps of building custom Copilot agent of your own.

First steps in building your agent

Before we begin, it’s worth emphasizing that you need a paid Microsoft 365 Copilot account to build your own agent that connects to your Microsoft data.

Any Microsoft 365 user (even those without the paid Copilot license) can create simple declarative agents using Copilot Chat. These agents do not access your Microsoft data (for example connecting to your email or SharePoint).

You can find the link “Create agent” on the left-hand side of the Copilot interface (for instance, accessed through microsoft365.com).

Clicking that link takes you to Copilot Studio, which looks eerily similar to ChatGPT’s interface for creating custom GPTs.

The left hand-side of the screen is where we build our agent, while the right-hand side is where we can preview it.

More specifically, on the left-hand side, you can build your agent via a conversation, at least as long as the “Describe” tab near the top of the screen has been selected. Simply describe how you’d like the agent to behave and Copilot Studio will start putting it together. If, however, you select the “Configure” tab, you can directly specify the instructions for your agent.

Significantly, if you proceed conversationally via the “Describe” approach, the name, description, instructions, and other aspects of the “Configure” approach will automatically be filled out.

Building an agent fit for an academic chair

One of the many aspects of my (Porter’s) job as co-chair of Drake’s mathematics and computer science department is to keep track of the independent studies supervised by my colleagues to ensure that they are adequately compensated for this work. Occasionally, it helps me to access information about the number of credit hours one of my colleagues has accumulated or how many independent studies were offered in a given semester. Although it’s not too much trouble to direct pull this information, it still takes time to extract it. Such a routine task is a great opportunity for a Copilot agent, so let’s give it a whirl!

Note here that since Microsoft Copilot has access to the files in my Teams folder associated with my work account, it can use the specific file I use to track independent studies to respond to my queries.

Checking in on the “Configure” tab, we see that updates have been made automatically.

Now that I have a first draft of my agent, I can try it out on the right-hand side of the Copilot Studio interface. I’ll see if it accurately reports the independent studies that I’ve supervised (I’ve anonymized the data for the sake of my former students):

So far, so good. At this point, I can continue the conversation to further refine the agent. I experimented with the agent to see if it keeps track of new changes to the underlying Excel file that it’s referencing. Here’s a fake entry I just added to the file:

Now I can check to see if the agent can report this new entry:

It worked! (I should probably delete this entry before proceeding further…)

Once I’m ready to finalize my agent, I can click “Create” in the upper-right corner of the Copilot Studio interface:

Note that the default is for the agent to only be accessible to me, but I can share the agent with others by clicking “Share”:

My options include sharing the agent with anyone in my organization or only with specific users in my organization (like my co-chair, who might find the agent to be helpful).

Once I create my agent, it can be found on the left-hand side bar where we got started:

It’s also worth noting that I can update my agent if I find that it needs further adjustments after having created it.

Clicking “Edit” just sends me back to Copilot Studio, where I can continue the conversation to improve my agent.

One final caution is that these agents are not perfect, as they can still make the kind of mistakes we’ve seen with large language models. So for high-stakes work, it’s important to double-check the work of your agent.

Want to learn more?

This has been just a brief overview of how to create your own Copilot agent. As you can imagine, there’s quite a bit more you can do with Copilot agents. To learn more, take a look at Microsoft’s overview of building agents. In the meantime, if you have access to Microsoft 365 Copilot, try building an agent of your own! Be sure to tell us how it went by emailing us at [email protected].