SBCO: Self-Supervised, Verifier-Grounded Harness Optimization For Planning Agents

SBCO optimizes planning agent performance using self-supervised, verifier-grounded harness optimization to bypass the alignment constraints of self-referential code editing. The method enables self-improvement in domains where task competence does not naturally overlap with code modification competence, reducing the computational overhead typically required by meta-agent self-modification or population-based approaches.