AI Agents vs Chatbots

A chatbot answers the next message. An agent takes a job and tries to finish it. Most people are buying the second thing and still living in the first.

That mix-up is expensive. You pay agent prices for a chat window. Or you ask a chatbot to run a week of research and wonder why the output dies in the thread.

What each one actually is

A chatbot is a conversation. You ask, it replies, you ask again. The next step is yours. The work stays in the thread unless you copy it out.

An agent is a system that can plan more than one step, use tools, and keep going until a goal is met or it has to stop. Anthropic draws the line this way: a workflow follows a path you already wrote. An agent decides its own next tool call (Building effective agents).

The research name for that loop is ReAct: reason, act, look at the result, then decide again (ReAct). If there is no loop and no tool, you have a chatbot. If the model can search, write a file, or come back with a draft you can open, you are closer to an agent.

When a chatbot is enough

Use a chatbot when the job is one turn, or a few turns you are happy to steer.

• You need a first draft of an email and you will rewrite it.
• You want a definition, a rewrite, or a second opinion on a paragraph you already have.
• The source is already in the prompt. Nothing else has to be fetched.
• You are exploring. You do not yet know what the deliverable is.

A chatbot is the right tool for thinking out loud. It is a bad place to store the result. If the only copy of the work is the last bubble in the thread, you still do not have a file.

When you actually need an agent

You need an agent when the next step is unknown, and the work has to leave the chat.

• The job takes more than one look-up. A research report that has to check sources, drop dead links, and come back as a document.
• Something has to change outside the model: a file, a spreadsheet, a ticket, a workspace.
• You will be away. The system should keep going, then stop at a checkpoint you can review.
• The output has to be handed to someone else. A client does not read your chat history.

Anthropic's rule is the useful one: start simple. Add an agent only when a single model call, or a short chat, cannot do the job (Building effective agents). Agents cost more, take longer, and fail in longer chains. That trade only pays when the job is open-ended and the finish line is a file.

What breaks if you pick wrong

Ask a chatbot to "do the research" and you get a confident summary with no trail. You cannot see what it opened. You cannot hand the thread to a colleague without a 20-minute recap. The work evaporates when the tab closes.

Ask an agent to "just chat" and you burn credits on a loop that did not need tools. It searches when a sentence would have done. It writes a 12-page pack for a two-line question. The extra machinery is not a feature if the job was one reply.

The failure mode that actually hurts teams is the first one: treating a chatbot like an employee, then discovering on Friday that nothing was saved.

Side by side

Three jobs, two tools

Research. "What is an AI research report?" is a chatbot question. "Read these six sources, drop anything older than 2024, and write a two-page brief I can send" is an agent job. The second one needs a plan, tools, and a document at the end. MMARW already treats that document as the unit of work, not the thread (The Artifact Standard).

Inbox. "Rewrite this reply so it is shorter" is a chatbot. "Triage this week's unread, draft the five that need a human, and leave me a list" is an agent. If you still paste every email into a chat window, you do not have an inbox agent. You have a slower chatbot.

A report. "Give me three bullets on this topic" is a chatbot. "Produce the client-ready file: question, method, findings, sources, what we did not check" is an agent. The test is simple. Can someone else open the output without you in the room?

How to choose in thirty seconds

1. Is the deliverable a sentence or a file?
2. Does the model need to look something up, or is everything already in the prompt?
3. Will you sit there for every turn, or should it run and come back?

Sentence, no lookup, you stay: chatbot. File, lookup, it comes back: agent. Mixed answers usually mean start with a chatbot, then promote the job once you know what "done" looks like.

If you want to feel the difference, do not debate the terms. Give the same brief to a chat window and to a workspace that can search, write, and save. The one that leaves you a URL is the agent.

Try that as a workspace on mmarw.com. Start free.

Sources

• Anthropic, Building effective agents
• Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models
• MMARW, What Are AI Agents?
• MMARW, The Artifact Standard