PhysEvo implements physical recursive self-improvement by decoupling a frozen model into a task agent and a meta-agent that diagnoses failures and revises tools. This framework enables joint-level control and reusable skills across 42 RoboDojo tasks without model-weight updates or separate action policy training.