Stop with the “AI agents vs agentic AI” debate

Welcome back to Building AI Agents, your biweekly guide to everything new in the AI agent field!

Who says LLMs don’t have a sense of humor???

In today’s issue…

• AI agents, agentic AI, and workflows: the differences

• OpenAI’s guide to building agents

• The professional network for agents

• Principles for building reliable LLM applications

…and more

🔍 SPOTLIGHT

Does being an agent make something agentic? Some people don’t seem to think so.

As “AI agents” have rocketed to prominence, an increasingly confused debate has emerged over what the term actually means. How independent does an LLM-based system need to be to be considered one? Is a chatbot an agent? Probably not, but what if it is given tools like web searching and memory? The dustup, naturally, has been exacerbated by the massive financial implications—everyone wants to call their LLM product an “agent.”

Personally, I think the best definition was proposed by LangChain CEO Harrison Chase: “An AI agent is a system that uses an LLM to determine the control flow of an application.” If a program just runs on rails, following the same pre-defined flow each time, it’s not an agent; if, however, the LLM can change the flow, then it is.

But what does it mean to “determine the control flow?” If ChatGPT does a web search to answer your question, it is choosing to activate some code that is part of its application—does this count, and therefore make it an agent?

My answer is that it does—sort of. But not to the same degree as a much more sophisticated multi-agent system in which a manager agent accepts a task, decomposes it into sub-tasks, and assigns them to several independent sub-agents equipped with a variety of callable tools, all without direct human input.

In both cases, the LLM is “determining the control flow” of its application. We’ll call this “agentic” behavior. But it is clearly doing so to a much greater degree in the latter case than the former. “Agentic AI”, therefore, is a spectrum, not a binary. This fits with Andrew Ng’s argument for adopting “agentic” as a compromise term to avoid pedantic debates over the exact dividing line between “agents” and non-agents. AI agents are clearly agentic, but not everything that has some agentic behavior is necessarily a fully-fledged agent.

There are some people—see here and here—who make the opposite case: that “AI agents” are simple LLM apps designed to automate a single, narrow task, while “agentic AI” describes systems that have a higher level of autonomy, but this seems plainly wrong—how could something be an “agent,” but not “agentic?” To be a “hero,” you obviously have to do something “heroic,” but not everyone who does something heroic is a hero. The opposite would be nonsensical.

What, though, do we call systems that clearly have some “agency”—the ability to use tools and/or reason toward their outputs, but that still largely carry out a series of pre-defined steps to accomplish a narrow task? Luckily, we already have a word for this: “workflows,” a term of art which long predates AI agents. Add LLMs to them, and we can call them “agentic workflows” if we want, but if we’re already talking about agentic AI, the first word is implied. So you set up an LLM to scrape LinkedIn listings and email you ones that match your experience—is that a full-blown “agent?” Debatable, but it’s definitely at least a little bit “agentic.”

In conclusion, agentic AI is a broad field encompassing many levels of autonomy—from largely rigid workflows with some LLM integration all the way up to the kind of humanoid artificial general intelligence companies like OpenAI are striving towards. Where exactly do you the draw the line to call something an AI agent? Well, that’s for you to decide.