What does agentic mean?

What it means

A normal chat answers once: you ask, it replies. An agentic system runs in a loop. It looks at the goal, picks an action, uses a tool such as reading a file, running a command or calling an API, looks at the result, and decides what to do next. It repeats until the task is finished or it gets stuck.

Coding tools like Claude Code and Codex work this way. Computer science has studied software agents for decades under the name intelligent agent. What changed is that a language model can now choose the next step from a goal written in plain English. The model can still be confidently wrong, as our page on transformer models explains, so the loop needs limits.

Why it matters to a business

• Narrow access: an agent can only damage what it can reach. Give it the least access that lets it do the job.
• Approval steps: a person should approve any step that is hard to undo and read the result before it goes live.
• Logs: you cannot undo an action you do not know about, so keep a record of every command the agent ran and every file it changed.
• Cost: every pass through the loop is billed model use, and a task that wanders makes many passes.

How this shows up in our work

Our engineers use coding agents in software work, and nothing an agent writes ships until a person has reviewed it. In AI and automation projects we agree with the client which systems an automated tool may read or change, and we keep records of what it did. Part of that conversation is which company data may go to an AI service at all. Agents sign in with keys, which we cover under AI API keys.

Learn more

• Intelligent agent on WikipediaThe long-standing definition of an agent in computing, for readers who want the background.
• Claude Code documentationOne agent tool and its permission model, as a working example of the idea.