A chatbot answers: you type a question and it replies. An AI agent acts: you give it a task, and it works out the steps and carries them out with tools such as your order system or email. Ask a webshop's chatbot about a return and it explains the policy. Ask an agent, and it can look up the order, create the return label and draft the reply.
Where is the line between the two?
Anthropic draws it in Trustworthy agents in practice (April 2026). A couple of years earlier, it writes, AI models were broadly available only as chatbots, "simple question-and-answer machines". An AI agent is a model that directs its own process and tool use to achieve what the user wants. The practical difference is a loop: the agent plans, acts, observes the result, adjusts and repeats until the task is done or it needs a person's input.
The word chatbot is much older than large language models. ELIZA (1966) spotted key words in what you typed and answered with prepared replies. Chatbots such as ChatGPT, Gemini and Claude generate their answers with a large language model (LLM). Earlier descriptions of agents, such as Lilian Weng's of June 2023, put planning, memory and tool use around such a model.
Is it a switch or a scale?
A scale. Anthropic's Building effective agents (December 2024) describes the steps between a plain chatbot and an agent:
- The augmented LLM: a language model with retrieval, tools and memory, which can write its own search queries and pick the right tool.
- Workflows: language models and tools follow code paths that a developer fixed in advance, such as a chain of prompts or a router that sends each question to the right handler.
- Agents: the model decides its own steps and tool calls, using feedback from its environment after each step.
Anthropic advises the simplest solution that works, which may mean no agentic system at all. Agentic systems often trade speed and cost for better results on the task.
Why does the difference matter for a business?
- An agent can do more. Tools let it fetch customer data, order history and help articles, and act on them: issue a refund or update a ticket. Anthropic calls customer support a natural fit for agents for that reason.
- An agent can also get more wrong. More autonomy means higher costs and errors that can pile up, so Anthropic recommends testing in a sandbox with guardrails.
- Control moves to permissions. In Claude's apps a user can set each action to always allowed, needs approval or blocked, for example reading the calendar freely but approving every invitation (human-in-the-loop).
Example
Anthropic's own example: asked to submit the receipts from a business trip, an agent plans the steps (transcribe each photo, take the amount and vendor, categorise the expense, submit it through the company's system) and works through them. If a hotel charge is flagged for exceeding the nightly cap, it may pause and ask whether to fetch the expense policy from the shared drive before trying again. Without tools, the same article notes, Claude can read the receipt but not file it.