Council Mode: Mitigating Hallucination and Bias in LLMs via Multi-Agent Consensus