AI agent vs. chatbot: what's the difference?
A chatbot answers your messages; an AI agent takes action. A chatbot produces text and stops, while an agent receives a goal, chooses tools such as web search, files, a terminal, or business software, works through several steps, and checks the result before handing control back. Both often use the same language model: the difference lies in the action loop and the permissions you grant.
The difference in one sentence: answering or acting
A chatbot works as a back-and-forth exchange: a message goes in, and a response comes out. If it gets something wrong, you have to notice and rephrase your request. It does not change anything outside the conversation.
An agent works in a loop: it reads the goal, chooses an action, carries it out with a tool, observes the result, and then decides what to do next, until it considers the goal complete. This “reason, act, observe” pattern was described as early as 2022 in the ReAct research paper. It now underpins almost all agents.
Anthropic, the company behind Claude, offers a useful distinction in its guide Building effective agents: a workflow follows a path defined in advance by a developer, while an agent decides its own path and which tools to use. There is therefore a whole spectrum between a chatbot and an autonomous agent.
Comparison table
| Criterion | Chatbot | AI agent |
|---|---|---|
| What it produces | Text or an image | Actions: edited files, sent emails, executed commands |
| Who decides the steps | You, one message at a time | The model, working in a loop |
| Tools | None, or a simple web search | Several, chosen at each step (tool calling, MCP servers) |
| Memory | The current conversation | Conversation plus files, notes, and task history (agent memory) |
| Task duration | A few seconds | A few minutes to several hours |
| Cost | One model call per exchange | One call per step, each sending the full context back to the model: costs add up quickly |
| Main risk | An incorrect answer | An incorrect action, or an action redirected by malicious content |
| Human oversight | At every message | Configurable: approval for every action, only sensitive actions, or none |
The cost row needs some explanation: an agent rereads its context, including instructions, history, and tool results, on every pass through the loop. A task with 30 steps therefore consumes far more than a single response. The LLM pricing calculator helps estimate that difference for your use case.
What makes a program an agent
There are four building blocks, reflected in the profiles in our AI agent directory:
- A language model that plans and makes decisions. It does not have to be the largest available: many coding agents run on medium-sized models, and the choice depends mainly on price and the reliability of tool calling. The model profiles show which models support it.
- Tools for reading and writing files, executing code, searching the web, controlling a browser, or calling a software application’s API. The MCP protocol provides a standard connection for these tools; our MCP server directory lists tens of thousands.
- A loop that runs actions in sequence and stops according to a rule: the goal is complete, the budget is spent, the maximum number of steps is reached, or human approval is needed.
- State: what the agent knows about the current task, such as open files, completed steps, and errors encountered. It sometimes also has persistent memory that carries over between sessions.
Remove the loop and tools, and you have a chatbot. Add them to a chatbot, and it becomes an agent.
Practical examples
The same request, two different behaviors. “Find September’s unpaid invoices and follow up with the customers.”
- A chatbot explains how to do it, suggests a follow-up email template, and stops there.
- An agent connected to your invoicing software and email lists the invoices, checks payments received, and drafts a follow-up for each customer. Depending on its configuration, it submits the drafts for approval or sends them.
The main types of agents in 2026:
- Coding agents read a repository, edit files, run tests, and make corrections until the tests pass. Examples include Claude Code, Codex, and Aider. The Claude Code vs. Codex comparison explains their differences.
- Self-hosted personal agents run on your computer or a small server, take instructions through a messaging app, and have access to your files and accounts. Examples include OpenClaw and Hermes Agent.
- Browser agents fill out forms and navigate websites for you, as the agent mode in ChatGPT does.
- Automation agents, built without code in n8n or Make, process emails, support tickets, or customer records according to rules.
Why the distinction matters: the risks change
With a chatbot, an error stays on the screen. With an agent, it becomes an action: a deleted file, a sent email, or a paid order. An agent that reads external content, such as web pages, emails, and documents, may also encounter hidden instructions that redirect its behavior. This is prompt injection.
Researcher Simon Willison describes the danger as a combination he calls the “lethal trifecta”: an agent with access to private data, exposure to untrusted content, and the ability to communicate externally can be tricked into leaking your data. An isolated chatbot has none of these three capabilities; a personal agent often has all three.
Three habits to adopt before moving from a chatbot to an agent:
- Grant only the minimum permissions required (least privilege).
- Require human approval for irreversible actions, such as payments, sending messages, or deleting files.
- Run the agent in an isolated environment when it executes code (sandbox).
The AI agent security assessment checks these points in a few minutes, and the guide to securing an AI agent explains them in detail.
Which should you choose?
| What you need | The right tool |
|---|---|
| Ask a question, draft text, summarize, or translate | A chatbot is enough and costs less |
| A repetitive task with the same steps every time | An automated workflow without an agent is more reliable and predictable |
| A task with multiple steps, where the path varies from case to case | An agent, with approval required for sensitive actions |
| A long task involving code or files | A coding agent or personal agent in an isolated environment |
OpenAI’s A practical guide to building agents makes the same recommendation: reserve agents for situations where fixed rules are insufficient, such as nuanced decisions, too many rules to manage, or unstructured data.
If you are deciding between agents, the Which AI agent should you choose? quiz asks five questions and recommends three options based on your needs.
Key takeaways
- A chatbot answers; an agent uses tools in a loop to work toward a goal.
- They can use the same model: the difference lies in the loop, the tools, and the permissions granted.
- An agent costs more per task and introduces different risks, including harmful actions and data leaks.
- Further reading: AI agents: definition, how they work, and practical examples and Getting started with an AI agent.
Frequently asked questions
Is ChatGPT an AI agent?
ChatGPT's standard chat interface works as a chatbot: you type a message, and it responds. ChatGPT's agent mode opens a virtual browser and terminal, chooses actions, and works through multiple steps, making it an agent. The same product can offer both.
Is an AI agent smarter than a chatbot?
Not necessarily: they often use the same language model. The difference is what the system allows the model to do through tools, an action loop, and memory, rather than a more capable model. A poorly configured agent using a good model makes more mistakes than a chatbot with clear boundaries.
Do you need to know how to code to use an AI agent?
No, if you use consumer agents, such as an assistant's agent mode, or no-code tools such as n8n or Make. You need at least some coding knowledge to install an open-source agent such as OpenClaw or build an agent with a framework. In every case, you need to understand what the agent is about to do before approving it.
Is a customer service chatbot an agent?
If it only answers questions from a knowledge base, it is a chatbot. If it can look up your order, issue a refund, or reschedule an appointment in the company's software, it takes action. That makes it an agent, with the risks that come with that access.