A roller coaster. This week has been a roller coaster.

It featured moments of fear, elation, and wonder crammed into five days.

From a Monday morning email about my recent LinkedIn posts and comments about AI to a DM asking to connect about consulting opportunities to a conversation at the end of the week, my emotions have whipsawed.

At the end of the ride, as I stepped off the call having said, confidently, I have built agents in ChatGPT and in Copilot, I didn’t know if we were talking about the same thing.

Did I get out over my skis, or was I on solid ground?

What is an agent?

This question nagged at me all weekend, so I resolved to figure it out.

Agent Defined

According to the Merriam-Webster dictionary, the noun “agent” has several definitions including:

• a  means or instrument by which a guiding intelligence achieves a result
• computer application designed to automate certain tasks (such as gathering information online)

That seems pretty straightforward. But it couldn’t be that simple, could it?

Turns out, no.

Three Opinions

I asked three LLMs the same two questions:

1. What is an AI agent?
2. Are Custom GPTs or Gemini Gems agents?

Because LLMs respond to prompts probabilistically, meaning they generate unique responses every time, all three responses were slightly different.

Leo, the privacy-first LLM available in the Brave browser, returned two clear answers:

“…a software system that can perceive its environment, make decisions, and take actions to achieve specific goals. “

“Custom GPTs (from OpenAI) and Gemini (from Google) are AI models that can be used to build AI agents, but they are not agents themselves….”

Although that wasn’t what I hoped to hear, I pressed on.

Gemini’s Answer

Google Gemini took a separate approach. It defined an agent exhaustively, and provided this excellent explainer video by Jeff Su in its response.

When I posed the second question, its answer was unequivocal:

“Yes, Gemini Gems are Google’s term for custom AI agents that you can create within the Gemini platform. They fit the definition of an AI agent because they are autonomous systems (built on the powerful Gemini model) that are given a specific role and instructions to consistently perform tasks on your behalf to achieve a defined goal.”

Wonderful. That was exactly the answer I wanted.

But, as I reminded a friend this week, when something seems too good to be true, it probably is.

ChatGPT’s Clarification

Its definition of an agent didn’t vary much from the prior two. However, because I use it to do the bulk of my AI work, because it knew what I was concerned about regarding the question, and because it is intimately familiar with Custom GPTs I’ve built, it drew a brighter, more distinctive line.

The Custom GPTs I made and use—one example here—are instruction-bound agents. They have specific, set goals and reason for a single purpose.

Another type of agent is an autonomous or multi-agent system. It differs from a Custom GPT because it works like an autonomous intern which knows its job, can collaborate with others as needed, works without direct supervision, and can complete its work with some independent decision making and initiative.

Which is where I’m not at. Yet.

Where From Here

I have a couple things to do.

First, I’ll circle back on that phone call and clarify what I meant by “agent.” What I said was accurate. I have built agents, but potentially not the type of an agent they had in mind.

And, more importantly, now that I have a fuller sense of what an agent could be, I can get started on transforming a current Custom GPT into a more autonomous and independent tool. I have a few ideas about where to start.

The key thing to keep in mind is that, no matter how many headlines you’ve read about AI recently, it’s still early days for the technology.

There’s little we can’t learn or understand more fully using AI. We just need to lean in and enjoy the rollercoaster ride.

Follow along, and I’ll share updates along the way.

Image: Photo by Charlotte Coneybeer on Unsplash