The subagent API improved earlier today, in v0.7.6.

There's currently the experimental gadgets-subagents package which showcases predeclared subagents.

Most coding harnesses also have the ability to, in one way or another, invent agent personas, conjuring agents out of thin air. For example, if I'm building a SPA, I can just go "put together a front end and a back end expert from available skills", or even just rely on the LLM's best guess without giving it any skills, and pray for the best.

I was wondering what that would look like in Agents SDK. My instinct is to re-expose the subagent API via a tool. And because it's likely that you'll need multiple ones, have some way to do it in bulk.

I've vibe-coded timdp#2 as my take on this for inspiration, but the ergonomics could probably be greatly improved.

#368 heavily relates to this: what if all the agents start streaming responses back to the main one? And by extension, what if they need to talk to each other, like Claude Code teams?

@threepointone Eager to hear your thoughts.