CERA-MoA (Co-Evolving Router with continually learning Agents for Mixture-of-Agents) introduces an iterative reinforcement learning framework designed to synchronize query routing with the evolving capabilities of LLM agents. Unlike traditional Mixture-of-Agents paradigms that treat routing and agent fine-tuning as disconnected processes, CERA-MoA enables the dynamic router and independent agent policies to co-evolve. To optimize efficiency, the framework utilizes a predictive familiarity estimator that leverages mid-layer hidden states to assess semantic competence across agents, bypassing the computational overhead of full rollouts. This estimation feeds into a cumulative-threshold adaptive routing mechanism, which selects a minimal, tailored subset of agents for each query to balance performance and resource consumption. Furthermore, the system promotes agent specialization by proactively allocating specific training samples to agents based on their measured competence levels. By facilitating this synergistic relationship between routing logic and agent training, CERA-MoA achieves superior performance compared to existing static-agent routing methods and fixed-workflow fine-tuning baselines across multiple domains.