AgensFlow: AI learns to route tasks in multi-agent systems, beating static pipelines

AgensFlow, developed by Nicole Koenigstein and published on arXiv, reimagines how multi-agent systems (MAS) built on LLMs are coordinated. Traditional approaches rely on static pipelines—fixed choices for which agent takes which role, which model to use, and how agents interact. These a priori decisions often fail under varying task regimes and operational constraints. AgensFlow instead frames coordination as an online policy-learning problem. It makes every coordination step—skill protocol, role assignment, model binding, interaction topology, retrieval, verification—observable and learnable from repeated trajectories. This allows the system to dynamically adapt routing decisions based on past performance, effectively learning the optimal orchestration strategy for each task type.

The framework was evaluated on two real-world corpora: distributed-systems incident tasks and security-advisory tasks. Results show three key findings. First, learned routing consistently reached higher-quality operating points than fixed pipeline baselines, especially on coordination-heavy tasks. Second, the 'skip:X' mechanism revealed that topology compression (skipping unnecessary steps) is a meaningful component of the substrate, improving efficiency without sacrificing quality. Third, warm-started policy graphs—initialized from prior runs—reduced exploration cost by up to 40% while maintaining plateau quality. The paper includes 4 figures and 4 tables, with open-source code and reproducible evaluations available. AgensFlow represents a shift from static MAS design to dynamic, auditable coordination that can be iteratively improved.