Research on clientelism emphasizes the use of brokers to mobilize voters. To utilize these agents efficiently, politicians must learn about brokers’ relative abilities and allocate scarce resources accordingly. Drawing upon a hand-coded dataset based on the archives of Gustavo Capanema, a powerful mid-twentieth-century congressman from Minas Gerais, Brazil, this paper offers the first direct evidence of such learning dynamics. The analysis concentrates on Brazil’s pre-secret ballot era, a time when measuring broker performance was particularly straightforward. Consistent with theories of political learning, the data demonstrate that resource flows to local machines were contingent on the deviation between actual and expected votes received in previous elections. Moreover, given politicians’ ability to discern mobilization capacity, payments to brokers were highly effective in bringing out the vote.