What is an AI coding agent?

Autocomplete, chat and agent: the difference

Autocomplete suggests the next line while you type. A chat answers a question and you paste the code wherever you want. An agent goes further: it takes a task, creates and edits files, runs the build, reads the error and tries again, over several steps, inside the repository.

Claude Code runs in the terminal and acts on the project folder. Cursor is an editor with the agent built in. GitHub Copilot lives in the editor and on GitHub. All three do the same underlying job: turning an instruction into code changes that you review.

What changes for the person building

You steer instead of typing: describe, review, test and correct course. That is the definition of vibe coding. The agent does not know what you want; the clarity of the task becomes the main skill, and reviewing what comes back is what separates a project that works from one that only looks like it does.

No tool replaces opening the project and using it. That habit is exactly what the IOOVC committee reproduces when judging.

Agents and the IOOVC

The olympiad requires no specific tool and asks for no proof of how the code was written. What is judged is the finished project, on the four criteria, with a score from 0 to 100.

How It Works (whether what you promised actually works, with attention to UI/UX, smoothness and usability) and polish (bugs, instability, accessibility and security problems) are where code generated without review usually loses points: the committee signs in with the test login and password and uses what was promised.

When this matters