“Agent” has been used to mean roughly ten different things this year, which is a shame, because the actual idea is small and worth having. A chatbot is reactive: you ask, it answers, it stops. An agent is given a goal and drives the steps itself — acting, seeing what came back, and choosing what to do next, over and over, until the goal is met or it hits something only you can decide.
What turns a model into an agent is having hands. A system that can only produce text can advise you; a system that can read a file, run a command, or search the web can go and find out. That’s tool use, and it is the whole mechanism. It also explains a behaviour people find surprising at first: agents recover from being wrong. Because each step ends with observing a real result rather than a guess, a bad first move produces an error the agent can read and route around — much the way a person would.
Most tools an agent reaches for are text-shaped, but not all. Computer use is the one that isn’t: Claude driving an actual screen, clicking through a real application and reporting what it saw. It’s the answer for work whose success you can only confirm by looking at it, and it’s the clearest illustration of how far the loop extends past a chat window.
The consequence of all this is supervision, not fear. Something that can take real actions on your machine should ask before the consequential ones, and it does — which is why permissions is a core topic rather than a settings footnote. And if your instinct is that this sounds like the process automation your company already bought, the comparison below is worth reading: the difference between replaying a recorded sequence and reading a situation is exactly why one breaks when the screen changes and the other needs reviewing like a new hire.
The words
- Agent
- An AI that doesn't just answer — it takes actions in a loop to reach a goal: read a file, decide the next step, run a command, check the result, repeat. Claude Code is an agent. A plain chatbot only talks; an agent does things on your behalf.
- Agentic AI Agentic workflows
- Agentic AI is an AI that works out the steps itself toward a goal — making a plan, taking actions with tools, checking the result, and adjusting — instead of doing the single thing you spelled out. A plain chat reply answers your one question and stops; an agentic system keeps going until the job is actually done, looping through act-and-check on its own. Claude Code is agentic: you give it a goal, and it reads files, runs commands, and fixes its own mistakes along the way.
- Tool use function calling
- Tool use is when an AI doesn't just talk but reaches for a tool to get something done — reading a file, running a command, searching the web. The model decides which tool fits, uses it, and reads the result back. It's what turns a chatbot into an
agent.
How it works in Claude
Commonly confused with
Two things people mix up here, and the difference that actually matters.