SBCO Optimizes Planning Agents with Self-Supervised Harness Feedback.

Traditional self-improving AI agents, like G\"odel Machines, often rely on self-referential code editing, which is effective for coding tasks but less applicable to other domains like planning. Adapting these methods for non-coding tasks typically involves computationally expensive population-based or self-modification searches. This new research introduces SBCO, a more efficient alternative for planning tasks with explicit constraints. SBCO operates as a self-supervised, verifier-grounded harness optimizer. Instead of self-referential modification, it learns a decomposed bank of verifiers and a harness policy through approximate block coordinate ascent. This allows the agent to improve its outputs using its own graded feedback, eliminating the need for human labels or a dynamic meta-agent. Experiments across two domains demonstrate that SBCO achieves performance comparable to or better than customized self-modifying baselines, while using 4 to 5.5 times less computational budget. This represents a significant step towards more efficient and scalable self-improving AI for planning applications.