AI agents, explained: what they do on their own, and where they stop
An agent is a model that can take actions, not just write answers. What that means in practice, what works today and what still needs a human.
A chatbot answers. An agent acts. Give it access to tools, such as a browser, your files or your calendar, and it can work through a multi-step task: look something up, decide what to do next, do it, and check the result.
What makes something an agent
Three ingredients: a capable model, a set of tools it is allowed to use, and a loop that lets it observe what happened and try again. Remove any one and you are back to a chatbot.
Where agents work well today
- Coding, where tests give quick, objective feedback.
- Research, where the job is to gather, compare and summarise.
- Operations chores, like triaging an inbox or updating a tracker.
- Anything repetitive with a clear definition of done.
Where they still need you
Agents struggle when the goal is vague, the stakes are high or the feedback is slow. They can also be misled by content they read, such as a web page with hidden instructions. That is why good agent products ask before spending money, sending messages or deleting anything.
Delegate the steps. Keep the judgement.
How to start
Pick one small, recurring task with a clear finish line. Give the agent only the access it needs, watch the first few runs closely, and widen its scope as it earns your trust.
