recent paper that gave coding agents data from social science experiments to see if they could reproduce them opencode scored the highest by doing more work if they set reasoning effort consistently the differences might just be from subagents