Every vendor with a generative AI product is calling it an “agent.” Most of them are not. Calling a workflow an agent is a marketing choice, and the difference matters for anyone trying to build with these tools or evaluate which ones are real.
Here is the line between workflows and agents, what each is good for, and how to tell which one you are actually looking at.
The shortest possible definition
A workflow is a fixed sequence of steps. A prompt goes in, the model processes it, tools may be called in a defined order, and a result comes out. The path is the same every time.
An agent is a system where the model itself decides what to do next, including which tools to call, when to stop, and how to recover when something does not work. The path is not fixed in advance. It emerges from the model’s reasoning.
That is the entire distinction. It is small in words and huge in practice.
What is being sold as an “agent” today
Most “AI agents” in 2026 are one of these:
- Workflows with an LLM at the front. The user asks a question, the model classifies it, a fixed pipeline handles it. Useful, but not an agent.
- Tool-using workflows. The model can call a known set of tools, but the order and conditions are coded. Still not an agent.
- Multi-step chains. A series of LLM calls, each feeding the next. Looks agentic from the outside, but the path is still baked in.
- Real agents. The model is given a goal and a set of tools, and it decides what to do. Currently brittle, expensive, and slow. The research goal, not the shipping product.
When a vendor says “agent,” ask which of those you are getting. The answer is usually the first three.
The four things that actually make an agent an agent
If you want to evaluate a system, here is what to look for. A real agent has all four.
- Planning. The model can break a high-level goal into sub-tasks without a human telling it the steps. If the breakdown is hard-coded, it is a workflow.
- Tool use with judgment. The model picks which tool to call, in what order, with what arguments. If a developer wrote the if-then-else tree, it is a workflow.
- Memory and state. The agent carries context across steps and revisits earlier decisions. Pure single-shot chains do not qualify.
- Recovery from failure. When a tool returns an error or an unexpected result, the agent adapts rather than crashing or stopping. Workflows stop or fall over.
If a system has one or two of these, it is on a spectrum. If it has none, calling it an agent is marketing.
Why the distinction matters
Because the failure modes are completely different.
Workflows fail predictably. If the prompt classifier is wrong, the workflow sends the request down the wrong path. You can debug this. You can add a test. You can ship a fix. The system does not surprise you.
Agents fail unpredictably. The model decided to do something a human would not. It might work. It might cost you a lot of money. It might take an action you cannot easily reverse. The system surprises you, and “surprise” is the wrong default for anything that touches production.
This is why most production AI systems in 2026 are workflows, not agents. Predictability wins. The cost of a confused agent is higher than the cost of a slightly dumb workflow.
Where agents actually make sense
There are real use cases, but they tend to share some properties:
- Sandboxed environments. The agent can act, but its actions are easy to roll back or ignore. A code agent in a throwaway branch. A research agent with read-only access.
- Expensive human effort being replaced. When the alternative is a human doing 40 hours of grind, even a slow and imperfect agent is a win.
- Reversible consequences. The agent can try things, see what works, and back out. Drafting emails, not sending them. Researching, not buying.
- High observability. Every action is logged, every step is inspectable, every decision is auditable. The team can see what the agent did and why.
If your use case does not meet at least two of those, you probably want a workflow with a good LLM at the front.
What to ask a vendor (or yourself)
When someone tells you their product is an AI agent, here is the question that matters most:
“Show me a run where the model did something the developer did not explicitly program.”
If the answer is silence, or a slide deck, you are looking at a workflow with good marketing. If the answer is a screen recording of the model improvising, you are looking at an agent. Both are useful. They are just not the same thing.