We present Adaptive Skill Coordination (ASC) - an approach for accomplishing long-horizon tasks (e.g., mobile pick-and-place, consisting of navigating to an object, picking it, navigating to another location, placing it, repeating). ASC consists of three components - (1) a library of basic visuomotor skills (navigation, pick, place), (2) a skill coordination policy that chooses which skills are appropriate to use when, and (3) a corrective policy that adapts pre-trained skills when out-of-distribution states are perceived. All components of ASC rely only on onboard visual and proprioceptive sensing, without access to privileged information like pre-built maps or precise object locations, easing real-world deployment. We train ASC in simulated indoor environments, and deploy it zero-shot in two novel real-world environments on the Boston Dynamics Spot robot. ASC achieves near-perfect performance at mobile pick-and-place, succeeding in 59/60 (98%) episodes, while sequentially executing skills succeeds in only 44/60 (73%) episodes. It is robust to hand-off errors, changes in the environment layout, dynamic obstacles (e.g., people), and unexpected disturbances, making it an ideal framework for complex, long-horizon tasks. Supplementary videos available at adaptiveskillcoordination.github.io.