The rapid growth of applications with diverse computational and network demands exacerbates challenges like uneven resource distribution and management complexity. The Computing Force Network (CFN) emerges to integrate the orchestration of computing and networking resources. Leveraging Software-Defined Networking (SDN), this paper proposes RL-CFR, a reinforcement learning-based multipath routing method that holistically considers both computational and network states. Within a Knowledge-Defined Networking (KDN) architecture, RL-CFR intelligently selects routes by evaluating service node computation status, flow latency, transmission rate, and packet loss. By utilizing SDN’s global control, RL-CFR precomputes and installs optimal paths. Experimental results demonstrate that RL-CFR significantly outperforms benchmarks in latency, link utilization, and packet loss, validating its effectiveness and practicality.