Currently, dify’s agents are relatively monolithic. It’s not possible to enable large models to autonomously decide which existing agents in dify to select and trigger, nor to select and invoke multiple
This is less about Dify and more about the underlying runtime.
Multi-agent systems only work well if you can dynamically load and switch models. Most setups assume everything is already running, which doesn’t scale well in practice.
dify不是需要多智能体协调，而是需要跨工作流的层级的编排。最理想的情况，单个工作流20个节点以内，承载稳定重复的功能，然后跨工作流编排实现多组功能的协同工作。现在必须把一个工作端到端地塞进一个单一工作流，构建复杂不说，稳定性和组合灵活性也不好控制。所以非常欢迎下个版本把工作流变成skill/插件节点进行支持。（但是状态机和运行超时是大坑，希望开发组能够妥善解决）