At its core, an AI agent is both extremely complicated and relatively simple. It’s a model or a chatbot that has access to the applications you and your team use every day, with some instructions on how you do things.

Now, I’m going to build a sales agent that handles sales reporting, lead qualification, and a bunch of other tasks for our sales team. And the principles we’ll cover are the same whether you’re building a meeting prep agent, an SEO agent, or any other type of agent that can start doing work for you.

So let’s go to Gumloop and create an agent. By default, it’s going to go ahead and select a model. You can pick from any of the available models out there. So if this is your first agent, I’d recommend keeping it to the default we have selected here.

Now, if I want this agent to manage reporting for me and our sales team, I have to give it access to the same tools I use and my team uses to generate a sales report. So let’s go ahead and add in HubSpot, which is our CRM, then we can give it Gmail for sending reports to the team, reading emails from prospects, and Slack, so we can send those same reports and maybe summarize conversations we’re having around customers in different Slack channels.

Now, Gumloop doesn’t just connect to applications like Gmail and Slack. You can also use external data sets, like Apollo for enrichment, or FireCrawl for web scraping. Now, if you’re building an SEO agent, you can add Webflow, Google Docs, and SEMrush to start helping you draft content. You can add in any application you would use for the task you want your agent to take care of.

Now, let’s already give it a try and send out a prompt. What does our pipeline look like? Now, the model receives the message and starts trying to figure out the best path. It’s going to look at what applications it has access to and decide on what it thinks the right next step is. Checking in here, it’s going ahead and grabbing information from HubSpot around our deals, which is what I was hoping it would do. And it returned that to me in a nice little table.

Let’s go a little further. Generate me a beautiful report of what our pipeline looks like and email it to me. And the agent is going to start writing me that report with what it’s gathered and send it to me so I can look at it in my emails. That’s amazing.

So now our agent has two parts. It’s using a model and it’s using the right applications, the ones we’ve given it access to. But it’s not personalized how we like things. That’s where instructions come in.

Instructions are included with every message to the agent, also known as a system prompt. They help define your agent’s behavior. We can give it a tone and some guidance. “You are a sales agent for the Gumloop team. Be direct, concise. Always respond to users in the same language they address you.”

You can even add some simple rules and boundaries. “When generating a report, always send it by email to the person requesting it and generate that same report that they can look at in the chat.”

Now if I say “generate me a pipeline report,” and even if I type it in French or in German, it’s going to understand our lightweight process because it’s always receiving the instructions first, and then executing the right steps.

And just with that, we’ve built an agent. It’s really that simple.

Now I want you to think back to your use case that you started this lesson with, what you want your agent to take care of for you, and create that agent. Connect the right applications, the ones you would use, and give it some instructions and then start chatting with it. You’ll be shocked at what it can already do.

And in the next lesson, we’ll cover how to share what you’ve built with your whole team by bringing it where you work, where your team works, in Slack and email.