Suggestion versus execution
An autocomplete-style assistant predicts the next few lines of code based on what's around your cursor. You read the suggestion, accept or reject it, and the tool's job ends there — it has no idea whether the code it just wrote actually compiles, passes tests, or fits the rest of the file three functions down. Every consequence of that suggestion is yours to discover.
Claude Code works differently because it operates in a loop, not a single prediction. Given a task, it reads the relevant files, forms a plan, takes an action — editing a file, running a command, searching the codebase — observes the result, and decides what to do next based on that result. If a test fails after an edit, it sees the failure output and can revise its approach without you re-prompting it. This read-act-observe cycle is what "agentic" means in practice: the tool is not just generating text that looks like code, it is pursuing a goal across multiple steps and adjusting when reality disagrees with its plan.