You run an agent called the GitLab runner. You configure it with a token from your Gitlab.com group or repo (depending on where you want the runner to be available). You tag the runner and reference it by tagging CI jobs you want to run on it. The runner polls for outstanding jobs and then executes them. It is extremely flexible in how it does this. In our case, the runner manages a fleet of executors using docker+machine. I believe this is similar to how Gitlab does it internally.

See https://docs.gitlab.com/runner/