I started reading the original ChatDev paper and was chuckling to myself on how intentionally opaque these academic papers can be. They introduce a set based notation to describe how multiple chat conversations can be strung together. It’s honestly a waste of brain cells to make such a simple idea so complex.

But, I did like the idea that the multi-agent chats were limited to two agents, instead of systems like AutoGen. I haven’t yet dived into claude-flow yet, but I wondered whether it was possible to have a claude code agent, or subagent, can spawan a subagent and have a long running two-way conversation with it.

I asked claude and it said:

Based on the documentation, the short answer is no, subagents cannot spawn other subagents. This is an explicit architectural constraint.

I wasn’t ready to give up, so I asked:

is it possible for a claude code agent or subagent to spawn a claude code process as a background process and interface with that process in a conversation? thus simulating multi-agent (or specifically two agent) topology. and, to search for whether anyone has tried already

And, it turns out that there is GitHub issue #4182, where users discovered that subagents can work around the “no nested subagents” limitation by shelling out.

But, much more interesting is at the very end of its response it said:

The Bottom Line

Your intuition was correct—spawning claude -p processes is a viable (if hacky) approach that people use today. But it turns out Anthropic has already built the proper infrastructure for this (TeammateTool with inbox-based messaging between agents), it’s just hidden behind feature flags. The community is actively unlocking and experimenting with it right now.

Wait, what is this TeammateTool? A bit of searching, it I found emergent-minds’ post on the topic. It looks like I missed a post on HN while I was in Vegas and it looks like Claude is already working on productizing multi-agent coding systems. Exciting.