If you've been following news about AI, you've probably come across the term "AI agent." You may have also heard about "agentic AI."
If you're like me, you might have wondered what was the difference between a GPT and an AI agent. I found a fantastic article on LinkedIn that spells out the difference between the two quite nicely. In short, a GPT is a customized generative AI tool that is trained on a lot of information so that when a person engages with it, it can generate responses based on the information. Think a customer support GPT. It could be trained on all of the support articles published for a product so that when someone pings the customer support widget on a site, it provides answers based on those articles. That's a GPT.
An AI agent is a collection of tasks that a person sets up that work based on a trigger the human sets.
I decided to play around with creating my own AI agent after reading more about them, and especially after continuing to read LinkedIn posts like, "I was able to set up an agent in a few minutes that did this cool thing for me, and I didn't have to code any of it."
Cool! No coding necessary, whip up agents quickly, start automating some of the things I've been doing manually. I'm in!
Step 1: Identify the use case
I wanted an agent that actually would save me time on a task. I originally thought of creating an agent that would brainstorm content ideas for me, but, with the help of my husband, realized that was the equivalent of automating a prompt that said "help me with 3 ideas on the topic of..." Not very time-saving.
Then I found myself thinking about all the places I went to look for news about what was happening in the world of market research technology.
BINGO! Use case identified.
Step 2: Identify the tools
It turns out there are tools that are called "no-code" and then the tools that require knowing Python or javascript. I actually did try going the Microsoft agent building route, using Copilot to guide me through the process, because GitHub has a built-in tool that will help troubleshoot the code.
Two hours into troubleshooting why the example agent wouldn't work, I figured out the file name it had given me to tell the agent to grab was wrong.
So I scrapped that, and a month later went to look for a no-code option. I found Zapier and Make.com, which both offered a free version.
I tried Make.com because their free version seemed more flexible than Zapier's, and if I wanted to go for a paid version, their pricing was lower than Zapier's.
Step 3: Build the agent
So, this is where I have to question every person who said they were able to create an agent in minutes. Maybe it's the fact that I was trying to create an agent that would collect news articles from multiple websites, rather than just connecting single apps to each other in a sequence of actions, but just figuring out how to pull the news articles into a Google Sheet with the URL and a summary of the article took me a full day. I was using Make.com's AI assistant, and I finally got things to work at the end of the day.
Or so I thought.
I let the system run, and looked at the results the next day. Every summary basically said, "In order for me to summarize this article, I'd suggest..."
So I scrapped that and tried again, this time pulling everything into a Google Doc. I ended up having to create separate workflows, one per news source, rather than have one giant workflow that pulled from every news source I entered (I'd collected 10) for the GenAI tool to summarize.
Once I had those workflows done, I found I needed to pay for Make.com in order to access the AI Agent ability. I'd spent so much time building all of the workflows (called Scenarios) that I decided to pay for at least a month to see if I found the agentic AI thing to be worth it.
I pulled all of my news-summarizing scenarios into the AI Agent builder, and then got incredibly confused, because I couldn't find where to create the module to tell the scenarios to start working.
An hour and a YouTube video later, I learned I needed to create a new Scenario that pulled in the "AI Agent." Essentially, the AI Agent collected all of the Scenarios together in one place, so that the final Scenario was essentially just the trigger, the collection of tasks, and the output.
Two days of no-coding later...
Maybe I just had a really steep learning curve others didn't with this process. Maybe I had a more complicated use case than others had for their AI agents. But this agent took a lot more trial, error, and tenacity than I had expected. Not only that, there were definitely some technical aspects that I needed to learn along the way so that I could map the correct fields from one module to the next.
Is it worth it?
YES. When you have the right use case for creating an AI agent, taking the time to build one is definitely a time-saver in the long-run!
Would I create an agent that wrote a survey for me, then connected to a panel provider's API to launch, and analyze the data once the number of respondents desired had been met?
NO. I still need to review the survey questions, at the very least. If it were a qualitative study, I would want to double check the analysis for hallucinations or just misrepresentations of data.
But does this have promise for further automation of various time-consuming tasks? Absolutely. I'm sure I haven't scratched the surface of what others have been able to build (though I side-eye the people who have built agents that will write and post content of any variety, because I don't think we should be delegating original content creation to any AI), but I'm already building another agent to pull more information from specific keyword searches on sites to then summarize. I'm excited to keep playing with this, and I'd love to hear what use cases you've thought of for agentic AI, or use cases for which you've created your own AI agents!
