The direction-of-arrival estimation in co-prime arrays is formulated as gridless compressive sensing where the atoms can be anywhere in a continuous domain. A new algorithm, successive-atom-merging cyclic coordinate descent for atomic norm minimization, is proposed. The algorithm performs atom update in subdomains of the continuous space relying on atom merging, exhibits fast convergence and has practically feasible computational complexity. It is demonstrated in simulations that the proposed method is superior to the joint sparsity reconstruction method (JLASSO) and the MUSIC method with spatial smoothing (SS-MUSIC) in terms of several criteria.