Coding agents are good at reasoning about goals in the digital world, but bringing them onto robots has been hard. Writing a policy in one shot fails because the model does not understand the robot or the physics well enough, and iterating directly on hardware is slow, costly and unsafe. SimEX uses simulation as the laboratory. In Stage 1 a coding agent runs open-ended probe-and-optimize cycles in a simulated sandbox, inventing its own tasks and building a reusable robot toolbox. In Stage 2 it adapts that toolbox to the real robot with only five physical trials, replaying each trial in simulation to diagnose the failure and screening several candidate repairs before spending the next trial. No demonstrations, no hand-written skills, and about ten minutes of robot time per task.
