What is an AI agent? Definitions may vary amid hype

The AI world might feel a bit like Hollywood at the moment: Everybody has an agent.

We’ve been told for months that agents are the next big thing in the generative AI revolution. And prognosticators say 2025 will be the year they become a reality. These (more) autonomous systems can act beyond the realm of a chatbot to plan and perform multi-step tasks.

But what exactly does that mean? Are code generation and text summarization—things that chatbots have been able to do since ChatGPT’s early days—not tasks or actions, too?

Many of the early examples of agents on the market range widely in function, from Salesforce’s customer service and campaign planning agents to Perplexity’s shopping agent and OpenAI’s new all-purpose Operator, which can “handle a wide variety of repetitive browser tasks such as filling out forms, ordering groceries, and even creating memes.”

With all the hype floating around, we wanted to attempt to determine a concrete definition of what constitutes an AI agent and how it’s different from other types of AI systems.

Tale as old as time

It turns out that this is not a new question. In a 1996 essay, a pair of pioneering AI researchers at the University of Memphis, Stan Franklin and Art Graesser, set out to define once and for all what separates an agent from a regular old computer program.

“Workers involved in agent research have offered a variety of definitions, each hoping to explicate his or her use of the word ‘agent,’” Franklin and Graesser wrote. “These definitions range from the simple to the lengthy and demanding. We suspect that each of them grew directly out of the set of examples of agents that the definer had in mind.”

They go on to collect a litany of existing definitions—such as having the “ability for autonomous execution” or “a persistent software entity dedicated to a specific purpose”—and reason through some philosophical questions, like the essence of agency and whether or not a sea squirt, for instance, achieves it.

The upshot of their survey is that agents must sense and act within an environment and pursue an agenda over time with a goal of affecting the future. This simple definition, as they note, is overly broad to the point of encompassing thermostats, bacteria, and humans at the extremes.

Nearly 30 years later, the dilemma they describe rings true, as themes echo throughout current definitions offered up by the likes of Salesforce, IBM, and Amazon: Agents should be able to interact with other tools and plan and execute workflows on their own, these companies say. Agents are also typically broken into different types by general functionality and motivating frameworks.

A simple reflex agent reacts to a sense of its environment with a pre-programmed action, like a thermostat, while a model-based reflex agent—IBM gives an example of a Roomba—can adjust based on its own model of its environment. Goal-based agents attempt to achieve an objective in the context of a modeled environment, and a utility-based agent attempts to achieve a goal while maximizing a certain value or utility.

“A generative AI system is considered an agent when it goes beyond creating output based on a prompt and begins to act on information to achieve a desired outcome,” Cognizant CTO of AI Babak Hodjat told Tech Brew in an email. “In other words, an agent uses its knowledge and reasoning, typically by employing an LLM, to make one or more calls to a set of tools it has been given.”

A missing link

Hodjat has been working on agents for decades, including early work on what would eventually become Siri, which was then conceived as a multi-agent system. What’s different now from that older research is that agents finally have sufficient computing power and data availability, as well as a means to use language and communicate with one another, he told us.

“For the first time, AI agents can interact using natural language and understand complex prompts,” Hodjat said. “This is why the current excitement around agentic AI is justified—while the concept isn’t new, we finally have the tools to make it work at scale.”

Hodjat said he hopes to eventually see tech companies move from thinking of agents as “one-stop shops” to bigger networks of agents communicating with one another in service of a bigger goal.

Category confusion

That’s a common vision of what agents will eventually look like, in theory. But the current buzz around the term has created a flood of different products that may have muddied the waters a bit around a concrete classification, according to Don Schuerman, CTO at Pegasystems.

“People are just turning an agent into anything that’s got an LLM somewhere inside it, and it does stuff,” Schuerman told Tech Brew. “And if it does stuff, and that stuff involves using a large language model, people are sticking the agent name on it right now.”

Pegasystems, which handles workflow automation and management, has been using the term “agents” for at least two decades, Schuerman said. These agents were usually autonomous background processes, he said. What’s changed now with LLMs is the wider scope of their autonomy.

“The idea of this as being a little self-contained piece of software that can run something on your behalf actually is carried forward into the way that we’re looking at it now,” Schuerman said. “They have the ability to both reason and plan and figure out what they’re going to do, and then, based on tools they have access to, they go and take action.”

Karen Panetta, a Tufts University electrical and computer engineering professor and IEEE fellow, said there’s a chance that agents might once again fade into the background as they become more commonplace in new technology. Panetta said it’s important to define now what agents are so that people might better grasp how these decisions are made.

“We’re gonna see more embedded agents,” Panetta said. “The general public won’t know that there’s an agent actually driving or making a decision about something for them. And I think that that’s one thing that I’m most concerned about.”