To address the dynamic workflow scheduling problem in fog computing under resource uncertainty (DWSPFC-RU), this paper formulates a mathematical model with the objective of minimizing total tardiness, and proposes a multi-tree niching genetic programming-assisted deep reinforcement learning (MTNGPDRL) method to solve the problem. First, the DWSPFC-RU is decomposed into routing decision and sequencing decision subproblems, and the routing agent and sequencing agent are then constructed. Second, a multi-tree niching genetic programming (MTNGP) algorithm based on the niche strategy is proposed to generate efficient and diverse routing and sequencing rules. On this basis, the rule learning capability of the MTNGP is integrated with the dueling double deep $Q $-network to construct a multi-agent deep reinforcement learning framework assisted by the MTNGP. According to the characteristics of the DWSPFC-RU, the state spaces and reward functions are designed for the two agents, and the action sets are constructed based on the rules generated by the MTNGP. Finally, simulation experiments and comparative analysis are conducted on test instances of different scales, and the results verify the effectiveness and robustness of the proposed method in solving the DWSPFC-RU.