What is an AI agent? Definition and examples

An AI agent is software that uses a language model to reach a goal by choosing its own next steps, calling tools and reacting to the results.

How it works

A classic chatbot takes one message and returns one reply. An agent runs a loop instead: it reads the goal, picks an action, looks at what happened and decides what to do next. Anthropic's engineering guide describes agents as "typically just LLMs using tools based on environmental feedback in a loop" and contrasts them with workflows, where the sequence of steps is fixed in code (Anthropic).

Three parts do most of the work:

• A model. Usually a large language model that reads the conversation and decides the next step.
• Tools. Functions the agent is allowed to call, such as "search the docs", "look up an order" or "book a meeting". Open standards such as the Model Context Protocol describe how tools and data are offered to agents (MCP docs).
• Guardrails. Rules about what the agent may do on its own, what needs a confirmation, and when it must hand the conversation to a person.

Why it matters for your website

Agents touch your website from two directions.

First, you can run one. An agent that answers visitors from your own content, asks two qualifying questions and books a meeting does in one conversation what used to take a form, an email thread and a calendar link.

Second, other people's agents visit you. Agentic browsers and AI assistants read pages, compare vendors and fill in forms on behalf of their users. A site whose prices, limits and next steps are easy to read and act on is easier for those agents to recommend.

The same Anthropic guide advises starting with the simplest system that works and adding autonomy only when the task needs it. For most websites that means narrow agents with a few well-defined tools rather than an open-ended assistant that can do anything.

Example

A visitor asks a software vendor's site: "Do you integrate with HubSpot, and can I see a demo this week?"

• A plain chatbot matches a FAQ entry and replies with a paragraph about integrations.
• An agent searches the docs, answers with a link to the HubSpot setup page, checks the sales calendar, offers three open slots, books the one the visitor picks and sends the new contact to the CRM. If the visitor asks about a custom contract term it cannot verify, it hands the conversation to a person instead of guessing.

Related terms

• Large language model (LLM): the model an agent reasons with.
• Model Context Protocol (MCP): an open standard for giving agents tools and data.
• WebMCP: a proposed way for web pages to offer tools to agents in the browser.
• Agentic browser: a browser that can act on pages for its user.
• Human handoff: passing a conversation from an agent to a person.