DynaFlow lets a main agent synthesize multi-agent workflows
Original title: DynaFlow: Learning Free-Form Multi-Agent Orchestration
With a fixed Qwen3.5-27B backbone, Forced mode reached 88.08% overall, 10.90 points above the strongest baseline.
DynaFlow adds free-form executable workflow synthesis to a main agent's action space so it can decide when to collaborate and write task-specific programs that coordinate fixed sub-agents. Programs support sequential or parallel execution, structured information exchange, and verification, with traces feeding later decisions. Only the main agent is trained, first by supervised fine-tuning on distilled trajectories and then on examples filtered for unsupported premises, biased verification, and information-flow defects. Across six benchmarks in math, scientific QA, and multi-hop retrieval, Forced mode scored 88.08% overall with the same Qwen3.5-27B backbone.
- Sub-agents stay fixed so orchestration learning is separate from executor improvement.
- Forced mode scored 88.08% overall accuracy across six benchmarks.
- That result beat the strongest evaluated baseline by 10.90 percentage points.
- The backbone was Qwen3.5-27B.
- Training used distilled trajectories, then a filtered refinement set.
- AI models
- Qwen3.5-27BDynaFlow
From the original article
1 min read at huggingface.co
Multi-agent systems often rely on predefined collaboration structures, limiting their ability to adapt to diverse task requirements. We introduce DynaFlow, a framework that integrates free-form executable workflow synthesis into the action space of the main agent.
By Ruhao Liu, Qi Li, Shizun Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.