Agents Lookup

Learn / Basics

What an AI agent is, and isn't

A plain definition of an AI agent, the three parts every agent has, and the tools that get the label without earning it.

Updated 2026-11-19. "(source N)" refers to the numbered list at the end.

The short definition

An AI agent is software that is handed a goal and works out the steps to reach it. It reads the situation, picks a tool, looks at what came back, and decides what to do next. It repeats that loop until the job is done or it gets stuck.

OpenAI's guide for builders puts it in one line: agents are systems that get tasks done for you, largely on their own (source 1). Anthropic draws the line in a similar place. In its terms, an agent is a system where the model steers its own process and its own tool use, rather than following a path a programmer wrote in advance (source 2).

So the key word is control. If the language model decides what happens next, you have an agent. If your code decides and the model only fills in a blank, you have something else.

The three parts of every agent

Strip any agent down and you find the same three pieces (source 1):

Two kinds of tools

Tools come in two flavours. Some fetch information, like reading a PDF or querying a database. Others take action, like sending an email or issuing a refund (source 1). The action tools are where the risk sits, so they are also where you will want the tightest rules.

Tools are also how agents plug into the rest of your software. The Model Context Protocol, or MCP, is an open standard that lets an agent connect to calendars, files, databases and business apps in a consistent way (source 5). We cover it in its own article.

What an agent is not

Plenty of products now carry the word agent. Many of them are not. OpenAI's guide is blunt here: if an app uses a language model but the model does not run the workflow, it is not an agent. Its examples are simple chatbots, one-shot prompts and sentiment classifiers (source 1).

Simple is often right

None of these are bad. Often they are the right answer. Anthropic's advice to builders is to start with the simplest thing that works and only add autonomy when it clearly pays off, because agents cost more and take longer to run (source 2). For simple, single judgments, see also our article on decision models.

What a good agent does that a script cannot

The case for an agent is a task with too many branches to write down. OpenAI lists three signs: decisions that need judgment and exceptions, rule sets that have grown too tangled to maintain, and work that leans on messy text like emails, contracts or claims (source 1).

Here is a concrete example. A customer emails: the blue jacket arrived in the wrong size, they want the large, and they are moving house next week. A script cannot handle that. An agent can read the email, look up the order, check stock for the large, see the new address, and draft an exchange. A careful setup then stops and asks a person before it ships anything.

Good agents also know when to stop. The guide says an agent should recognise when a task is finished, correct itself when a step fails, and hand control back to a person when it cannot go on (source 1).

Where agents are today

Agents are real and in use, but they are not magic. In LangChain's survey of more than 1,300 practitioners, 57.3% said they had agents running in production. About a third named quality as their biggest problem, ahead of speed and cost (source 3).

Results also vary a lot by task. Stanford's 2026 AI Index reports that agent success on Terminal-Bench, a test of command-line work, rose from 20% to 77.3% in a year. The same report notes robots still succeed on only 12% of real household tasks (source 4). Software tasks with clear checks are where agents do best today.

The practical upshot for a small business: treat an agent like a capable new hire on probation. Give it a narrow job, the minimum access it needs, and a person to check its work until it has earned more trust.

Quick test: is it an agent?

Ask three questions about any product that claims to be an agent.

Reading the result

If the answer to all three is yes, you are looking at an agent. If not, you may be looking at a workflow or a chatbot with a new label. That may still be exactly what you need.

Sources

Written by Agents Lookup in our own words. Facts come from these pages: