Recent advances in software testing allow automatic derivation of tests that reach almost any desired point in the source code. There is, however, a fundamental problem with the general idea of targeting one distinct test coverage goal at a time: Coverage goals are neither independent of each other, nor is test generation for any particular coverage goal guaranteed to succeed. We present EVOSUITE, a search-based approach that optimizes whole test suites towards satisfying a coverage criterion, rather than generating distinct test cases directed towards distinct coverage goals. Evaluated on open source libraries and an industrial case study for a total of 1,752 classes, we show that EVOSUITE achieved up to 188 times the coverage of a traditional approach targeting single branches, with up to 62% smaller test suites.