LOSI: IMPROVING MULTI-AGENT R | Pangram Labs