What Is an AI Agent? Chatbots vs Assistants vs Agents

“Agent” is one of the most used — and most stretched — words in AI. Some products call any chatbot an agent. Others reserve the word for systems that work on their own for hours. This guide gives you a clear definition and a practical sense of when agents are worth it.

Chatbot, assistant, agent

Feature	Chatbot	Assistant	Agent
What it does	Answers messages	Answers and helps with tasks using some tools	Pursues a goal by deciding and taking actions in a loop
Who decides the steps	Fixed script or one reply	You, turn by turn	The model, within limits you set
Typical tools	None or a FAQ	Web search, file reading, image generation	Code execution, files, browser, apps and APIs
Example	A website FAQ bot	Asking an AI to summarise a PDF	“Fix the failing tests in this project and open a pull request”

The key difference is who controls the sequence of steps. In a chat, you do: you ask, read, and ask again. An agent receives a goal and decides by itself what to do next, many times over, until it is done or needs your input.

How the agent loop works

Under the hood an agent is surprisingly simple. Anthropic’s widely cited guide describes agents as “LLMs using tools based on environmental feedback in a loop”:

1. Goal. You describe what you want and any constraints.
2. Decide. The model chooses an action — read a file, search, run a command, call an API.
3. Act. The application executes that action (the model never runs anything itself; it requests tool use through function calling).
4. Observe. The result goes back to the model: the file contents, the error message, the search results.
5. Repeat until the goal is met, a limit is reached, or the agent asks you for a decision.

The quality of an agent therefore depends on three things: how capable the model is, how good its tools are, and how clear its goal and limits are.

Types of agents you will meet

Coding agents. The most mature category. Tools like Claude Code, OpenAI’s Codex, Cursor’s agents and GitHub Copilot’s coding agent can read a codebase, edit several files, run tests and iterate on errors. Code is a good fit because results can be checked automatically — tests pass or fail.

Browser and computer-use agents. These operate a web browser or a desktop like a person would: clicking, typing and reading the screen. They are useful for tasks on sites without an API, but slower and more error-prone than direct integrations.

Research agents. They search many sources, read them and write a report with citations. Often several sub-agents read in parallel while a lead agent combines the findings.

Workflow agents. Inside automation tools like n8n, Make or Zapier, an agent step can choose between business tools — check an order, look up a customer, search the help centre — before drafting a reply.

Many agents connect to apps through MCP, an open standard that lets one connector work across different AI applications.

When an agent is worth it

Agents trade cost and speed for flexibility. They are a good choice when:

• the task has many steps that you cannot predict in advance;
• the outcome is valuable enough to justify more time and model usage;
• mistakes can be detected and fixed — by tests, review or undo.

If the steps are always the same, a fixed workflow is cheaper, faster and more reliable. If a single well-crafted prompt can do the job, use that. A useful rule from practitioners: start with the simplest thing that works, and add autonomy only when it clearly improves results.

Where agents fail

• Compounding errors. A wrong assumption early on can send the agent down a long, costly path.
• Overconfidence. Agents may report success when a task is only partly done. Check the result, not the summary.
• Prompt injection. Content the agent reads — a web page, an email — can contain hidden instructions. An agent with broad permissions could be tricked into acting on them.
• Runaway cost. Long loops read and write many tokens.

Using agents safely

• Give the minimum permissions needed: read-only where possible, a sandbox or separate account for experiments.
• Require approval for irreversible actions: sending messages, payments, deleting data, publishing.
• Set limits on time, steps or budget.
• Review the output — diffs, drafts, logs — before accepting it.
• Connect only tools and MCP servers you trust.

To see how agents are designed in practice, continue with AI agent design patterns.