What is an AI agent?

An AI agent is a software system that uses a language model to work through a task step by step without being told each step. It plans, calls tools, reads what comes back and decides what to do next. So unlike a chatbot, it does not stop at answering questions: it takes action in the systems it is connected to.

How does an AI agent work?

An AI agent runs in a loop. The model reads the task and its context, picks a tool, gets the result back and decides whether to continue or stop. The loop ends when the task is done or when a limit you have set kicks in.

Anthropic describes the basic building block as a language model augmented with retrieval, tools and memory. In practice, an agent has five parts:

• Model: the language model that plans and makes decisions.
• Tools: functions and APIs such as document search, an ERP lookup or drafting an email. Open standards like the Model Context Protocol define a common way for applications to connect to tools and data sources.
• Instructions: the rules for what the agent should and should not do.
• Memory and knowledge: the history of the current task plus the company knowledge it is allowed to read.
• Limits: permissions, human approval steps and a maximum number of iterations.

OpenAI's guide to building agents names model, tools and instructions as the three core components.

AI agent vs. workflow vs. chatbot

The difference is who decides the sequence of steps. In a workflow, code fixes the steps and the model fills in individual gaps. In an agent, the model itself decides which steps and tools are needed. A chatbot only replies and never acts.

Anthropic advises starting with the simplest solution that works and adding complexity only when it is needed. In its framing, agentic systems trade latency and cost for better results on open-ended tasks. Workflows are more predictable; agents are more flexible.

A practical example

Picture a service agent handling a fault report about a machine. It searches manuals and past tickets, looks up the serial number in the ERP if it needs to, and drafts a reply that cites its sources. A person reviews the draft and sends it.

Which documents it reads and which systems it queries depends on the request, and the agent works that out for itself. That is what sets it apart from a fixed process. For a closer look at this kind of setup, see the page on the AI agent for technical service.

What to watch out for

An AI agent is only as dependable as the limits around it. Before it goes live, permissions, approvals and logging need to be settled.

• Permissions: the agent sees only what the person it works for is allowed to see.
• Approval: anything that leaves the company, such as an email to a supplier, is confirmed by a human first.
• Traceability: every step and every source is logged.
• Data residency: the model and the data can also run entirely on your own infrastructure.

For an overview of where agents fit in mid-sized companies, see AI agents.