The difference from autocomplete is the loop: the agent plans, acts, reads the result and tries again until the task is done or it gets stuck. Examples are Claude Code, GitHub Copilot in agent mode and Cursor. An agent is only as good as what it can read — a README that is accurate, tests it can run and a pipeline that objects count for more than which model was chosen.
That puts new demands on the codebase, which is why we built the AI readiness check and wrote about why your next platform user is an agent (in Swedish). The contract between people and agent is usually written in an AGENTS.md.