AI agent

What it is

An AI agent is a program that pursues a goal over several steps on its own. It uses a language model to decide what to do next, carries out that step with a tool, looks at the result and continues until the goal is reached or it has to stop.¹²

How it works

Goal. A person or another system sets the task, for example “find three suppliers and request quotes”.
Plan. The model breaks the goal into steps.
Act. It calls a tool: searches the web, opens a file, fills in a form, sends a message. Tool access is increasingly standardised, for example through the Model Context Protocol.⁴
Observe. It reads what came back.
Check and repeat. It decides whether the step worked and what comes next.

Agent, chatbot, automation

	Chatbot	Classic automation	AI agent
Decides next step	No	No, fixed rules	Yes
Uses tools	Rarely	Yes, pre-defined	Yes, chooses itself
Handles surprises	Answers only	Stops or fails	Tries another way
Predictability	High	Very high	Lower

What businesses use it for

Research and preparing documents from many sources.
Sales support: finding contacts, drafting first messages for a person to approve.
Back office: sorting emails, updating records, reconciling lists.
Software development: writing, testing and fixing code.

Risks

Mistakes add up. A small error early on can lead to many wrong steps.
Prompt injection. Text on a web page or in an email can contain hidden instructions the agent follows.³
Too many rights. An agent that may pay, delete or send can do real damage.
Cost. Every step uses the model; long tasks become expensive.

Introducing agents safely

Give it only the rights the task needs.
A person approves anything that cannot be undone: payments, deletions, messages to customers.
Set spending and time limits.
Log every action so it can be traced.
Start with one narrow task and measure the result.

Key terms

Tool call: The moment the agent uses an external function, for example a search or an email.
MCP: Model Context Protocol, an open standard for connecting models to tools and data.⁴
Human in the loop: A person approves certain steps before they happen.
Prompt injection: Hidden instructions in content the agent reads.³

Latest research

New papers and reports that mention this term, found by our daily source scan. One line per source, quoted as published.