DynaFlow learns free-form multi-agent orchestration
A research paper by Ruhao Liu, Qi Li, Shizun Wang, and Xinchao Wang introduces DynaFlow, which adds executable workflow synthesis to a main agent’s action space. The agent can decide when to collaborate and generate task-specific programs for sequential or parallel sub-agent execution, structured information exchange, and verification; the paper reports 88.08% overall accuracy in Forced mode across six benchmarks using a Qwen3.5-27B backbone.
The paper presents a method for learning adaptable multi-agent coordination while keeping the sub-agents fixed, separating orchestration learning from improvements to the underlying executors.
Computer Science > Multiagent Systems
arXiv:2610.34467 (cs)
[Submitted on 25 Oct 2026]
Title:DynaFlow: Learning Free-Form Multi-Agent Orchestration
Authors:Ruhao Liu, Qi Li, Shizun Wang, Xinchao Wang
View a PDF of the paper titled DynaFlow: Learning Free-Form Multi-Agent Orchestration, by Ruhao Liu and 3 other authors