Have you tried using an additional agent to verify the outputs? It seems that can help if the supervising agent has a small context demand on it. (ie. run this command, make sure it returns 0, invoke main coding agent with error message if it doesn't)