Current Mixture-of-Agents (MoA) frameworks treat query routing and agent fine‑tuning as independent steps, which hinders adaptation to agents that improve during post‑training. CERA‑MoA introduces a co‑evolutionary approach where the router and continually learning agents are updated iteratively, allowing routing strategies to track evolving agent capabilities. This joint optimization promotes synergistic, data‑driven specialization among agents in the mixture.