(P)rior(D)yna(F)low: A Priori Dynamic Workflow Construction via Multi-Agent Collaboration

Recent studies indicate well-designed LLM-coordinated workflows outperform single models in task-solving. Though more works focus on autonomous workflow construction, most rely only on historical experience, limiting efficiency and adaptability. We propose an a priori dynamic framework for automated workflow construction: it uses Q-table learning to optimize decision-making (utilizing historical experience) and lets agents evaluate task progress for a priori decisions on the next agent, selecting suitable structures. Cold-start, early stopping and pruning are added to boost efficiency. Experiments on four benchmarks validate its feasibility: it outperforms state-of-the-art baselines by 4.1% on average and cuts construction/inference costs to 30.68%–48.31% of existing methods.