Skip to main content
An experiment is your research container: a title, hypothesis, source, and place in a workflow. An artifact is the source you upload for that experiment. An agent version is the immutable identity created from a finalized artifact. The standard CLI workflow centers on a single agent.py source file. The benchmark API can also upload a multi-file folder for competitions that define a bundle entrypoint. A new version is created whenever a new snapshot is published; later edits to your local files do not change a completed evaluation or submission. This separation lets you compare exactly what ran, return to an earlier version, and submit the intended evidence-backed candidate.