What people mean by 'AI agents'
The jump from a chatbot that answers to a system that takes actions, and why that jump deserves more care.
Jamie Owen Updated 2 Jun 2026
A chatbot answers. An agent acts. The difference is that an agent is given tools (the ability to search the web, run code, send an email, update a record) plus a goal, and then decides for itself which steps will reach that goal. The model is the brain and the tools are its hands.
Those tools usually arrive through MCP servers, the standard most assistants now use to reach outside themselves.
Why it’s useful
For multi-step work, an agent can run the whole loop: read your calendar, draft the replies, file the notes, and only come back when it’s stuck. For routine tasks that used to mean a person clicking through several apps, that is a real saving.
Why it needs more care
The moment a model can do things, a mistake has consequences beyond a wrong sentence. An agent that misreads an instruction can email the wrong person or delete the wrong file. Agents are also a fresh target for prompt injection, where instructions hidden in the content they read hijack what they do.
Give an agent the narrowest set of tools and permissions that lets it do the job, and a human checkpoint before anything irreversible.
The sensible posture today: use agents for low-stakes, reversible work, and keep a person in the loop wherever an action is hard to undo. If your organisation runs on Microsoft 365, agents will most likely arrive through Copilot, which has a page of its own.