DynaFlow: Free-form multi-agent orchestration framework
In brief
- Main agent generates executable workflows to coordinate sub-agents without predefined structures.
- Training focuses solely on the main agent; sub-agents remain fixed.
- Results: +10.9 points accuracy across six benchmarks with the same base model.
Researchers introduce DynaFlow, a framework enabling a main agent to generate executable, adaptive workflows for coordinating sub-agents. According to the announcement, the agent autonomously decides when collaboration is needed and creates task-specific programs handling sequential or parallel execution, structured information exchange, and verification. Execution results guide subsequent actions, improving adaptation during reasoning. The paper reports tests on six benchmarks (mathematical reasoning, scientific QA, multi-hop retrieval) showing a 10.9 percentage point gain with the same Qwen3.5-27B model, reaching 88.08% accuracy in forced mode.
The approach separates orchestration learning from executor capability improvement by training only the main agent via supervised fine-tuning and data filtering. Workflow analysis shows more explicit use, verification, and refinement of upstream outputs post-training, highlighting cross-stage information flow as a key dimension.
This summary gives you the essentials. The full article, its figures and quotes are on Hugging Face Daily Papers.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error