Supporting deep agents and/or automatically generate sub-agents without predefined flow

[Wonder1905]

Are planning to support deep agents, or other methodology to spwan subagents without predefined graph or workflow?
similar to how Claude code/ opencode works?

Thanks!

[jeffreysijuntan]

Hi, we are indeed working over this direction. Supporting the agent itself should not be too hard. But do you have any specific tasks/dataset you have in mind to train the agent?

[Wonder1905]

Its less on which dataset, but more for saving context, for example letting a subagent read large html context, to save context exploding from main