I assume the 'random' strategy was the one on the left? It didn't ever extend the line - just the pole. Not a fair test?
Still working on the OpenAI plumbing though, so still a ways before I get to play with all the different research agents available.