Chatbot vs AI agent vs AI operating system

Open almost any AI product page this year and it calls itself an agent. The chatbot you already use has an agent mode, and older products have been relabeled to match, a habit Gartner calls agent washing. If you are trying to work out which of these will actually do work for you, the labels will not help. What separates a chatbot, an AI agent and an AI operating system is how much of the work lands back on you after the tool is done.

The short version

A chatbot returns text and hands every next step back to you, while an AI agent decides its own next step and uses tools to carry a task out.
A chat window says nothing about what sits behind it, which is why ChatGPT behaves as a chatbot in plain chat and as an agent in agent mode.
Gartner calls the relabeling of chatbots, assistants and automation tools as agents "agent washing" and estimates that about 130 of the thousands of agentic AI vendors are real.
An AI agent that runs as an app on an operating system built for a person still depends on that person for its context, its logins and the checking of its work.
An AI operating system holds the agent's identity, memory, accounts, permissions and record of actions, and those jobs exist whether you run one agent or fifty.

The short answer

A chatbot answers you in text and hands the work back. An AI agent takes a goal and works through the steps itself, choosing tools as it goes. An AI operating system is what an agent runs on. It holds the agent's identity, memory, accounts, permissions and a record of what it did, so you stop holding the agent together.

What is a chatbot?

A chatbot takes text in and gives text back. It can explain and draft, and then the next step is yours. Redis put the boundary in one line in its May 2026 comparison of the two: the chatbot "can tell you how to get a refund, not process one." Slack's guide to the agentic OS says the same from the other side: "Generative AI chatbots answer questions and create content, but they stop there."

A chat window also tells you very little about what is behind it. IBM's explainer on AI agents says chatbots "are a modality," meaning the chat box is how you talk to the system, and treats what the system can do as a separate question. IBM describes nonagentic chatbots as "ones without available tools, memory or reasoning." The same text box can sit in front of a plain model or in front of something that edits your files, so judge a tool by what happens after you press enter.

What makes something an AI agent?

An AI agent decides its own next step. Anthropic's engineering guide Building effective agents draws a clear line between the two kinds of system. Workflows, in its words, "are systems where LLMs and tools are orchestrated through predefined code paths." Agents "are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." The same guide says agents "are typically just LLMs using tools based on environmental feedback in a loop." It was published in December 2024 and now carries a note that much of the tooling it describes has changed, but the distinction has held up.

Redis describes that loop as one where the agent "reasons about a goal, calls external tools, observes the results, re-plans" and keeps going until the task is done. Tools by themselves are not the test. IBM's comparison of agents and assistants says it directly: "The ability to call on tools by itself does not make an LLM an agent." A chatbot with a web search plugin still waits for you after every answer. What changes the category is the model choosing its next move and carrying on without you steering each one.

That is also where the stakes go up. Redis notes that when a chatbot gets something wrong, "the blast radius is limited to incorrect text," while agents "compound this risk by taking irreversible actions."

Where do AI assistants fit?

Assistants sit between the two, and the line is about who keeps the task moving. IBM's comparison puts it this way: "AI assistants are reactive, performing tasks at your request." An assistant needs your direction through the task. An agent keeps working on its own after the opening prompt. Asking for a summary of your inbox is assistant work, and handing over a goal and getting back finished bookings is agent work.

Is ChatGPT a chatbot or an AI agent?

It is both, depending on the mode. Plain ChatGPT chat is chatbot behavior: you ask, it answers, and whatever comes next is up to you. In July 2025 OpenAI launched ChatGPT agent, which you switch on by selecting "agent mode" from the tools menu. The launch announcement says it carries out tasks "using its own virtual computer" and that "ChatGPT requests permission before taking actions of consequence." OpenAI now marks that post as outdated, so read it as a description of the launch and not as the current product spec.

For this comparison, look at where the agent works. In agent mode it runs on a virtual computer OpenAI provides for the task, with a browser and a terminal, and reports back into your conversation. Agents that work through an ordinary desktop instead, clicking and typing the way a person would, are covered in what is a computer use agent. In both cases the agent is a program working inside an environment someone else set up for it.

Why does everything call itself an agent now?

Because vendors are relabeling older products to ride the hype, and Gartner has a name for it. In a June 2025 press release, Gartner described "agent washing" as "the rebranding of existing products, such as AI assistants, robotic process automation (RPA) and chatbots, without substantial agentic capabilities." The same release says "Gartner estimates only about 130 of the thousands of agentic AI vendors are real," and 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.

You do not need a vendor list to sort this out. Two checks do most of the work. One is whether the tool decides its own next step or waits for you after every answer. The other is what it still remembers when it stops, and who has to pick the work back up. A tool that waits for you after every answer is a chatbot with a new label. A tool that acts on its own but forgets everything when it stops is a real agent that still leans on you.

Where does an AI operating system fit?

An AI operating system sits below the agent. Windows and macOS are operating systems built for a person, and an AI operating system is the same class of software built for an agent as its user. We define the term properly in what is an AI operating system.

Slack's guide frames an agentic OS as the layer that manages AI agents and how they work with people, data and tools, where a traditional operating system manages the machine and runs applications. Systems researchers are working on the same problem from the operating system side. In Agent Operating Systems (AOS), Ankur Sharma and Deep Shah write: "While agents can be implemented as user-space applications today, their execution characteristics stress OS boundaries in scheduling, memory and state management, security, observability, and governance." The AgenticOS workshop, held with SOSP 2026 in Prague on September 29, says traditional OS abstractions "were never designed for dynamic, semantically rich, adaptive agent workloads."

Do you need an AI operating system for one agent?

Business explainers mostly say no. They treat the agentic OS as a layer for coordinating many agents across a company. Aryan Agarwal of Gravity, in a September 2026 explainer, sets a personal threshold of ten or more agents in production, alongside conditions about teams and systems of record, and writes, "Under roughly ten agents it is overhead." The same piece quotes Requesty: "A single agent is an application."

That answer fits the question it is asking, which is how to supervise a fleet. The literal operating system answers a different question, which is where the agent lives. One persistent agent still needs an identity that survives a restart, memory that outlasts a session, logins nobody pastes into a prompt, permissions that something other than the prompt enforces, and a record of what it did. Those are operating system jobs whether there is one agent or fifty.

Run that single agent as an app on an operating system built for you, and those jobs fall to you. You re-explain the project each morning because the agent forgets between sessions, which we cover in why AI agents forget everything between sessions. You hand it credentials and decide how far it can reach, which is the subject of how much access an AI agent should have. Then you read back through its work to see what it actually did. We make the case for pairing each agent with its own operating system in one operating system, one persistent agent.