Builds are 100% stateless within the CI cluster and rely on tracking/publishing just about anything you can imagine or plug as a resource; out-of-the-box inclusions being docker images (registry), S3 (e.g. minio, AWS) and git.

It's a bit strange to reason about in the beginning but I have yet to try anything that comes remotely close to Concourse when it comes to orchestrating across multiple sources of input/output and non-linear pipelines.