No More Stale Feedback: Co-Evolving Critics for Open-World Agent Learning