Code completion, and the snippet you ask for in a chat, are still suggestions: you paste them in, you run them. A coding agent goes one step further — it opens the files, rewrites them, runs the test, reads the error message and tries again. The difference is not in the model, it is in what it has access to.
Its ability is measured by one well-defined number: how long a human task it finishes with a 50% success rate. For the best models today that is around five hours of human work. Fifty percent, though, means precisely that it gets every second one wrong — so your first task should not be somewhere that hurts.
The most common misconception is that a good agent removes the need for review. It is the other way round: the more it writes, the more there is to read. Reliability does not come from the model, it comes from having something the agent can run for itself — a test, a type check, a linter — because without that feedback it is working blind.
Choosing that first task, and the safety net around it, is what Your First Task with a Coding Agent walks you through. Directly related terms: AI agent, AGENTS.md and vibe coding.