Ivan Malopinsky

Defining agents

Defining “AI agents” is not easy these days. It’s a technical term that rapidly became a buzzword with many different meanings. Rather than insist on a specific definition, I think it’s useful to see how it’s actually used today.

“Agent” as a term has a long history in artificial intelligence, going back to the 1970s. The current wave of LLM-based agents started in 2023, building on decades of prior work. The hope is that by using large language models, we can build intelligent agents that can easily adapt to different situations, which is very difficult with previous approaches.

Considering all the ways agents are defined, I come back to three definitions I like:

The first is the business definition, useful for non-technical audiences, that emphasizes the role and capabilities of AI agents. The second is the software definition, using language familiar to technical people, emphasizing the nature of agents and how they differ from typical programs. The last one is the current technical definition - you’d have to know what all the terms mean in context, but it’s an excellent and punchy summary.

In practice, “AI agents” have been used in much broader ways:

The laziest marketers slap “agent” on top of whatever they were selling before just to ride the hype train. The promise of autonomous workers driven by powerful AI is the core of their pitch, and the details don’t matter.

Also, bizarrely, some leaders and companies brag about using thousands if not billions of AI agents, which I think points to how sloppy their definitions are. The number of agents is almost irrelevant: it’s what they do and how they do it that’s important.

After thinking about it for a while, I realized something about the way people use “AI agents.” Whenever it’s used in an imprecise way, there are either missing words or missing context. The words “AI” and “agent” can mean different things entirely depending on what’s being discussed.

For example, a chat bot is not, by itself, an “AI agent.” It’s neither “AI” nor an “agent.” However, what’s being sold is a replacement for an existing job, like a customer service agent (or a travel / insurance / real estate agent). Seen this way, the chat bot is an AI replacement for a customer service agent. It’s not necessary for the “agent” part to meet any technical definition since it’s a job title, and so these customer support chat bots tend to be basic software without a lot of AI or autonomy at their core.

Coding assistants are an interesting case because they’re already aimed at a technical market. Even so, they usually lack autonomy (unless given some by the user) and are typically used in a collaborative setting to create something, not necessarily to execute something. They could be seen as fancy website or app builders, but they’re not quite AI and not quite virtual workers. They are evolving, however, and could be the platform for “true” AI agents in the future.

Workflow automation vendors have really jumped on the “AI agent” bandwagon, shamelessly branding their “if this then that” and “take A and use B to turn it into C” workflows as real legitimate “AI agents.” There is, however, no agency because there’s no autonomy and no uncertainty. If the world changes, the workflows have to be changed, too. That said, I see workflows as a building block for systems using AI agents, and a critical one.

I’m sure there will be more examples, since everything is an AI agent now. Maybe at some point we’ll have a more useful term that hasn’t been diluted beyond utility. In the meantime, I’ll stick to my three favorite definitions.