Rethinking the value of multi-agent work-flow: a strong single agent baseline