A system that proposes and evaluates changes to its own operating components.
Self-improvement may change prompts, skills, routing, memory rules, tools, or harness configuration based on measured outcomes. Safe improvement requires held-out evaluations, versioning, rollback, and human governance of the objective itself.