"Agentic" is the 2025 word that has lost all meaning.
It is in pitch decks, startup positioning, research papers, and product marketing. It describes everything from a wrapper around ChatGPT that can call three APIs, to systems that genuinely plan and replan across multi-hour tasks. When a word applies to everything, it means nothing.
Here is a sharper division.
What is actually agentic
A system is meaningfully agentic if it does all of the following.
- Plans. Produces a sequence of steps before executing, not step-by-step reactively.
- Replans. Updates the plan when execution reveals something it didn't anticipate.
- Remembers. Carries state across steps that the model itself wouldn't remember from its context window alone.
- Chooses. Decides which tool to use and when, rather than being told which tool to call at each step.
Systems that do all four are rare. The ones I've seen that genuinely do it count in the low tens.
What's being called agentic but isn't
Most of what's shipping as "agentic AI" in 2025 is a script.
- A chain of prompts that route to hand-specified handlers. Not agentic. That's a decision tree with a language model inside.
- A tool-use loop that calls the same model with the same prompt until it produces valid JSON. Not agentic. That's a retry loop.
- A system that lets the model call tools but always in a fixed sequence. Not agentic. That's a script with RAG in it.
- A system that calls itself recursively without planning. Not agentic. That's a recursive call with extra steps.
These systems work. Some of them work well. None of them are doing what the word "agent" was meant to describe.
Why the distinction matters
If you're deploying a script, you can reason about its failure modes. Scripts have bounded output. You can test them. You can cap their behavior. When a script fails, it fails in predictable ways.
If you're deploying a real agent, you cannot do any of that. Its failure modes are novel. Its behavior is bounded only by the tools it has access to and the implicit constraints of its prompt. Agents fail by doing unexpected things. That's nearly the definition.
Teams that call their systems agentic and deploy them as if they were agents are fine, because their systems aren't actually agents. Teams that build real agents without realizing what they're signing up for are the ones getting in trouble.
The category that's getting dangerous
Somewhere in the middle are systems that are almost agents. They plan sometimes. They replan occasionally. They hold state across a few steps. They choose tools within a small menu.
These are the dangerous ones. Dangerous not because they are powerful. Dangerous because they are unpredictable enough to surprise you and not capable enough to recover from the surprise.
The clearest symptom: a system where the builders cannot confidently predict, given an input, what the output sequence will be. Not the exact text. The sequence. If that's unpredictable, the system is more agentic than its builders are treating it. That's the gap where real production accidents happen.
What to say instead
If you have a chain of prompts, call it a pipeline.
If you have a tool-use loop, call it a tool-use loop.
If you have something that plans and replans dynamically, call it an agent, and be honest about the additional surface area that comes with it.
Precise words do a specific job. "Agentic" currently does no job. Using it marks a product as following the hype curve. Using precise language marks a product as known to its own builders.
The technology is interesting. The word has stopped being.