Robot learning has made substantial progress by scaling demonstration data and model capacity, yet deploying a new robot or task still requires a dedicated teleoperation corpus, and the resulting policies, bounded by the demonstrations they imitate, struggle to generalize and self-improve. We introduce RAX (Robot Agent eXploration), an open-source cognitive architecture for robot agents whose core organizing principle is scaling exploration: robot agents should improve by gathering their own online experience and optimizing their programs. In RAX, a general-purpose coding agent programs the robot directly, without action demonstrations, hand-designed action primitives, or value function guidance. It explores online to discover the decisions that drive task success and automates them into modular robot programs that integrate perception, planning, and control. The underlying language model parameters remain fixed; instead, learning updates a persistent workspace of programs, notes, and verified skills. As exploration scales, the agent constructs task-relevant 3D perception, plans actions by searching robot joint poses in 3D, and generates fully automated controllers with closed-loop visual and force feedback. Using RAX, we conduct the first large-scale, multi-embodiment, systematic study of scaling exploration for robot agents, spanning fine-grained precision and long-horizon tasks on tabletop single-arm and bimanual manipulators, dexterous hands, mobile household robots, and humanoids, in simulation and in the real world. Performance on unseen scenes grows steadily as exploration scales. In simulation, RAX sets a new state of the art by a wide margin, surpassing the strongest prior models by +37.4 on average. On real robots, RAX learns complex memory-dependent, articulated, precise, deformable, and long-horizon mobile manipulation tasks through autonomous exploration alone, and verified skills accumulated across sessions let later sessions build on earlier ones. We release the full architecture of RAX as a generalist robot-agent platform, on which a new robot can be brought up through an interface alone, without data collection or policy training.