DynaFlow Moves Multi-Agent Systems from Fixed Pipelines to Dynamic Orchestration

Introduction

The central challenge in multi-agent systems is not simply whether more models can be called. It is deciding when collaboration is worthwhile, how subtasks should be assigned, and whether intermediate results can genuinely influence later reasoning. Many existing systems use fixed roles, topologies, or workflow templates. Such designs are easy to control, but they can be poorly matched to tasks with changing complexity and information dependencies.

DynaFlow takes a different approach. It places free-form executable workflow synthesis inside the action space of the main agent. The main agent can assess the current problem, decide whether delegation is necessary, and generate a task-specific program that coordinates multiple sub-agents.

Key ideas

• Collaboration on demand: The main agent may answer directly when a task is simple, while more difficult problems can trigger decomposition, retrieval, or cross-checking.
• Flexible execution patterns: Generated workflows can run sub-agents sequentially or in parallel and pass intermediate results through structured exchanges.
• Verification as part of execution: Outputs are not merely collected at the end; later stages can inspect, correct, and refine upstream results.
• Adaptation during reasoning: Execution outcomes and traces are fed back into subsequent decisions, allowing the collaboration plan to evolve as reasoning progresses.
• Separated training objectives: Supervised fine-tuning is applied only to the main agent. Sub-agents remain fixed, so the training target is orchestration rather than general executor improvement.

The training pipeline begins with supervised fine-tuning on distilled trajectories. It then refines the model with examples filtered for unsupported premises, biased verification, and information-flow defects. This focus is important: DynaFlow evaluates not only whether a workflow runs, but also whether evidence travels correctly across stages and whether verification actually contributes to the answer.

Results and implications

Across six benchmarks covering mathematical reasoning, scientific question answering, and multi-hop retrieval, DynaFlow reaches 88.08% overall accuracy in Forced mode. That is 10.90 percentage points higher than the strongest evaluated baseline with the same Qwen3.5-27B backbone. Workflow analysis reports more explicit use, verification, and refinement of upstream outputs after training.

The broader significance is a shift in how multi-agent systems are optimized. Instead of simply adding more agents, the emphasis moves toward designing useful information paths between them. An effective collaborative system needs intermediate outputs to be traceable, checkable, and reusable. Dynamic orchestration also creates unresolved trade-offs: the main agent may delegate unnecessarily, or skip collaboration when a workflow would have helped. The discussion around the work points to counterfactual evaluation as a way to measure both costs separately.

DynaFlow therefore represents a move from static multi-agent architecture toward task-level program generation. It does not attempt to improve every agent at once. Instead, it teaches a central agent to inspect the task, organize execution, and use feedback. That division offers a practical direction for building more adaptive and interpretable agent collaboration.

Source: Hugging Face Daily Papers