Let’s build our first agent, our first Gummie. Now I want to create an agent that can help our support team navigate when they should give a discount to a user. Our team should be able to drop in a Zendesk ticket number in Slack, and that’ll kick off the approval process fully managed by an agent.
So let’s go to Gumloop and create a new agent and call it Support Bot Assistant. Now, quick overview of the layout here. We’ve got tools my agent has access to—those are the applications it can use, or the Gumloop flows it can run. Then the instructions, which is how it should behave and when it should use different tools. And finally, a space to test our Gummie as if we’re talking to it in Slack.
So let’s start building our agent. The starting point is Zendesk, so I’ll add that as a tool. Let’s test it out by saying return details for a specific ticket. And we see it retrieves all of its details.
Now, the way I want this flow to work is that this Gummie should validate the ticket we’re talking about to make sure we’re on the same page. So to make sure that’s the case, I’ll add the information to the system prompt—something like “you’ll receive Zendesk ticket information, begin by returning the ticket details and confirming with the user that it’s the correct ticket.”
The next step in our flow is that eligibility for discount depends on a bunch of factors scattered across different systems. So let’s add a flow that goes to BigQuery, Salesforce and retrieves a profile of the user. Then we can update the instructions on when it should use it.
The final element of our agent is reading our policy on discounts, which lives in Notion, and assessing whether the user is eligible. We’ll add that connection to Notion, but we notice we’re giving it access to all Notion pages. That’s probably gonna overwhelm our poor little Gummie. So we’ll update the instructions with more details—specifying the exact Notion page URL with the discount eligibility criteria.
And just like that, we’ve built an agent. You give it a Zendesk ticket number, it’s gonna confirm with you whether it’s the correct ticket, then kick off user profile research through a flow, read the Notion doc with all the discount information, and assess whether the ticket qualifies. At any moment, you can interrupt or ask other questions about the policy.
What do you need to remember? Developing good agents is part art, part science. There’s a balance between improving the instructions, giving it too much or too little information, and adding the right tools. View your agents as a constant work in progress that you improve over time as more people use them.
In the next lesson, we’re gonna bring this Gummie into Slack.
