Hmm.. agent cannot do self-supervised learning without actually doing it. The trick is to keep it in a sandbox.