Some multilateral agencies implement aid projects in a broad range of sectors, with aid disbursements showing a strong overlap with those of bilateral donors. The question then arises of why do bilateral donors delegate sizable shares of their aid to non-specialized agencies for implementation? This paper develops a game theoretic model to explain this puzzle. Donors delegate aid implementation to the multilateral agency (ML) to strengthen the policy selectivity of aid, incentivizing policy improvements in recipient countries, in turn improving aid's development effectiveness. Bilateral donors are better off delegating aid to ML even when they are purely altruistic but disagree on how aid should be distributed across recipients. Key for our result to hold is that ML searches some middle ground among disagreeing donors. Aid selectivity-in terms of both policy and poverty-emerges endogenously and is credible, as it is the solution to ML's optimization problem. Moreover, the model shows that if one sufficiently large donor is policy selective in its aid allocations, there is no need for other donors to be policy selective. The World Bank's aid program for lower-income countries, the International Development Administration (IDA), is shown to fit the assumptions and predictions of the model. Specifically, IDA is a dominant donor in most of its recipient countries and is much more policy and poverty selective than bilateral aid. Donors view it as a public good, and contribution more to it when bilateral aid is less selective. Potential threats to IDA's role as a dominant, policy-selective donor include the emergence of nontraditional donors, changes in voting shares, and traditional donors' increasing use of earmarked contributions.