The Red Queen Gödel Machine: agents that co-evolve with their own evaluators

arXiv·medium signal

arXiv:2606.26294 (June 26) describes a self-improving agent architecture where the agent and the evaluator that scores it co-evolve, aiming to avoid the stagnation of optimizing against a fixed, gameable reward. It's a research-stage take on the self-improvement loop that underlies autonomous harnesses. Relevant for anyone building systems meant to get better run-over-run without a frozen benchmark to overfit.