How does a coding agent's scope differ from inline code completion?

A **code-completion model** predicts the next tokens given whatever is in your editor right now. It sees the open file, maybe a few neighbours, and guesses the continuation at your cursor. It is fast and useful, but it runs once and stops. It never executes the code it wrote and never finds out whether the suggestion was correct. A coding agent is the system around the model, not the model itself. It is given tools: read a file, write a file, run a shell command, run the tests. The agent uses these tools in a loop. It reads the relevant files, makes an edit, runs the test suite, and reads the output that comes back. That output is the key. When a test fails, the error message and stack trace come back into the agent's context as an observation. The agent then reasons about the failure and decides the next edit. It repeats read, edit, run, repair until the tests pass or it runs out of budget. So the agent owns the whole edit, test, debug loop. Completion owns one cursor position. The agent can verify its own work; completion cannot.

A **code-completion model** runs a single forward pass over a static context, usually the open file plus a little surrounding code, and predicts the next tokens. It cannot run the code it suggests, cannot read test output, and cannot react to its previous edit. A coding agent is an LLM placed in a loop with tools. It owns the full edit, test, and debug cycle. Each turn it reads files, writes an edit, runs the suite, and captures the result. The loop looks like this: $$\text{edit}_t \rightarrow \text{run tests} \rightarrow \text{observe output} \rightarrow \text{repair}_{t+1}$$ The defining difference is execution feedback. Test failures, stack traces, and compiler errors return into the agent's context as observations. The next reasoning step is grounded in what the code actually did, not in what the model predicted. This is exactly the workflow SWE-bench measures. An agent gets a real GitHub issue and a repository. It must navigate the code, edit it, and produce a patch that makes the hidden tests pass. Devin, the Cursor agent, Aider, and Claude Code all implement this loop. Completion tools like Copilot inline suggestions do not.