I have been thinking again about the definition of AI Agents over the weekend.
When I wrote AI Agents Demystified in 2024, I defined an AI Agent through four core elements: Sensor, Decision, Action, and Memory. That framework was not random. It came from the classical sense-act loop in AI: an agent senses the environment, makes decisions, takes action, and uses memory to improve over time.
Two years later, the meaning of “AI Agent” has become much more diluted.
Today, many so-called AI Agents are really LLM-powered workflows: a predefined chain of steps, sometimes with tool calls, sometimes with a chat interface, and usually with a very impressive demo video. They can be useful. They can save time. They can improve productivity.
But are they actual agents?
I do not think this is just a naming debate. Decision-making is not a small detail of an AI Agent. It is the whole point. Throughout history, humans have invented tools to execute our intent. Most tools are designed to produce predictable outcomes. AI Agents are different because they introduce something fundamentally new: a tool that can make decisions within a defined scope.
That is why the word matters.
If “agent” simply means any system that includes an LLM and runs through a few steps, we lose the word for the genuinely new thing. We turn a paradigm shift into a product label. Good for marketing, bad for understanding.
Anthropic draws the line clearly: workflows follow predefined code paths, while agents allow the LLM to dynamically direct its own process and tool use. In other words, the question is not whether an LLM is involved. The question is who controls the flow: the code, or the model?
That dividing line is simple. The market made it blurry.
Andrew Ng’s “agenticness” spectrum is a useful framing because real-world systems are not binary. Some systems are more agentic than others. But the spectrum also gave the industry permission to call almost anything on the low end an agent. Ng himself later said that the “agentic” sticker is now being slapped on everything by marketers. The industry now has a word for this: agent-washing. It means rebranding basic workflows, scripted automations, RPA, or chat interfaces as AI Agents without substantial agentic capabilities.
This is not harmless. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls.
Definitions have stakes. Loose definitions create failed projects.
That said, I would also refine my own 2024 framing. I may have underweighted how much practical agent deployment would happen at the low end of the spectrum. In reality, many useful systems today are workflow-heavy, with only thin slices of real autonomy at the edges. That is where the technology is more reliable, easier to control, and easier to ship.
So my updated view is this: Decision-making is still the defining trait of an agent. But in practice, most deployed systems exist on a spectrum.
For builders, the better question is not only “Is this an agent?” The better question is: How much real decision authority are we willing to give this system, and over what scope?
That question is much more useful.
If agents are just workflows, the conversation is mostly about productivity and ROI.
If agents are decision-makers, the conversation becomes much bigger. It becomes about trust, oversight, delegation, accountability, and failure recovery. It is about what humans remain responsible for, what we are willing to hand off, and how we know when something has gone wrong.
Those are very different conversations. The harder one is the one we need to have.