In　order　to　characterize　the　agent,　there　is　a　presupposition　that　each　agent　has　its　set　of　propositions.　The　model　is　based　upon　our　prior　observation,　and　satisfies　many　properties.　Furthermore,　we　describe　concretely　the　desire　of　agent　by　twin-subset　semantics　of　desire　operator.　If　friendship　for　learning　of　two　agents　have　been　made,　one　agent　has　the　authorization　to　increase　his　set　of　propositions　rationally,　namely,　learning　from　the　other　one.　The　paper　aims　to　make　one　agent　collaborates　rationally　with　the　other　one,　and　implement　BDI　logic　in　multi-agent　system.