While Large Language Models (LLMs) show promise in alleviating mental health resource shortages, existing approaches struggle with complex emotional dynamics and maintaining theory-grounded counseling trajectories. In particular, they fail to adapt interaction rhythms across counseling phases, often resulting in excessive exploration or premature interventions. To address these limitations, we propose AuroraCounsel, an agent-based counseling framework that models professional counseling workflows. The framework integrates an emotion-sensing agent for affect recognition, a reflection agent for structured psychological analysis, and a guidance agent for response generation. Through emotion-driven dynamic routing, AuroraCounsel selectively activates reflective reasoning only when negative emotions are detected, enabling accurate problem tracing while avoiding unnecessary computation. Building on this architecture, we further align counseling behaviors by optimizing the guidance agent with Grouped Relative Policy Optimization (GRPO), allowing it to regulate questioning and supportive strategies across dialogue phases and overcome the limitations of supervised fine-tuning in capturing clinical counseling rhythms, such as those prescribed by REBT. Evaluations using LLM-based evaluators and expert assessments by professional counselors demonstrate consistent improvements over existing approaches in professionalism, empathy, and interaction appropriateness.