What is an AI agent

An AI agent is a program that, as well as answering you, can do things on its own to carry out a job: open files, search the internet, send emails or run commands on your computer, deciding the steps itself. The difference with a chat is simple: the chat tells you what to do and the agent does it. That's why it's more useful and also more delicate.

Inside, an agent combines three things: a language model (the "brain" that understands and writes text), a set of tools it has access to, and a loop in which it thinks, acts, looks at the result and decides the next step. What's confusing is that "agent" has become a buzzword and gets slapped on any chatbot; one that only replies with canned phrases isn't an agent. Nor do they all have the same freedom: some ask for permission at every step and others do their own thing, which is exactly where the scares begin.

I work with agents every day, and I recently started trying out OpenClaw, an open-source agent you install on your own machine and control from messaging apps. My first impressions are in "I've started playing with OpenClaw". And what you need to watch out for when you give them real work, from permissions to costs, you'll find in "The problems of working with agents".