A multi-agent system can fail not because its agents lack capability, but because their handoffs are fixed before the task is understood. DynaFlow makes those handoffs part of the main agent’s reasoning: it writes task-specific workflows, then uses execution traces to decide what to do next.
Key Findings
- The main agent can coordinate fixed sub-agents through sequential or parallel calls, structured information exchange, and verification. This makes orchestration—not executor retraining—the central learned capability.
- Training uses distilled trajectories and filters examples for unsupported premises, biased verification, and information-flow defects. The emphasis on information flow matters because a later agent must actually use and check upstream work, not merely receive it.
- Across six benchmarks in mathematical reasoning, scientific question answering, and multi-hop retrieval, the paper reports 88.08% overall accuracy in Forced mode, 10.90 percentage points above the strongest evaluated baseline using the same Qwen3.5-27B backbone.
Who It’s For
Great fit for researchers studying adaptive agent coordination, workflow synthesis, and how intermediate results affect multi-stage reasoning. Look elsewhere if you need evidence of production latency, cost, or autonomous real-world deployment: the reported results focus on benchmark accuracy, and the design assumes a fixed set of sub-agents.
Method in Context
Rather than selecting from a predefined collaboration template, the main agent generates executable coordination logic and can adapt after seeing results. Keeping sub-agents fixed helps isolate the value of orchestration learning, though it also means the reported gains do not show that the framework improves the agents doing the underlying work.