This doesn’t seem to work when the harness feeds images and asks the agent to do things in the real world. It fails to devise ways to keep track of its progress and fails to utilize its tools effectively.