AI Agent
An AI system that plans and carries out multi-step tasks using tools, and why good agents have limits and human approval.
AI agent: an AI system that plans and carries out a task in several steps toward a goal, often by using tools such as search, email, files or a code editor, instead of only answering a single question.
In plain English
A chatbot answers. An agent acts. You give it a goal, such as “sort my support emails and draft replies”, and it works through the steps, deciding what to do next and calling tools along the way.
What an agent usually has
- A goal and instructions (often a system prompt).
- Tools it may use, such as search or app connections, often through MCP.
- Memory or context for what it has done so far.
- Rules and limits, including when it must stop and ask a person.
Why it matters
Agents can save time on multi-step work, but because they can act, mistakes have consequences. Good agents have minimum access, clear never-do rules and human approval for actions that send, publish, delete, spend money or affect other people.
Common confusion
- Agent does not mean fully autonomous. Most useful agents work with a person approving key actions.
- An agent is not a single tool. It is a way of using a model with instructions and tools.
- Text in emails and web pages can try to instruct an agent. It must treat that as data, not commands.
Related terms
System prompt, MCP, LLM.
Learn more
Read How to Write a Safe AI Agent Brief and see examples such as the Inbox Triage Agent.