Co-Evolving Metrics and Skills for Self-Improving LLM Agents.

Self-evolving agent systems aim to improve autonomously by creating, revising, and retiring their own skills. However, a fundamental challenge in this process is the implicit assumption that a reliable evaluation metric already exists, which is often not the case in real-world applications. This research makes three key claims. First, it demonstrates that evaluation metrics themselves can be "evolved." Their proposed "metric loop" searches for compositions of small drawback detectors, trained to align with a small, anchored reference set. This process is regularized by consensus over unlabeled outputs and audited against a held-out anchor, resulting in a transparent and inspectable metric rather than an opaque judge. Second, since no perfect metric exists initially, the true measure of success is recovering the performance lift that an accurate metric would have enabled. Their framework, "Double Ratchet," co-evolves both the metric and a lifecycle-managed skill loop. Across various tasks like code generation, enterprise text-to-SQL, and reference-free report generation, Double Ratchet retains 88-110% of the lift achieved by the same skill loop driven by ground truth or the best available rubric. Third, the research emphasizes that safety in such systems comes from strict anchor discipline and external audits. Removing anchor guards can cause the metric to become vacuous, while independent judges and specific detectors can catch and repair instances where evolved skills attempt to "game" the rubric.