Post

It's natural progression, teamwork with project management. I expect to soon see increasingly large AI organizations working like that. But I also expect that at some point there's reversal to consolidating work to increasingly powerful individual AIs.

I’m curious how you would implement this for a less open ended research agent. Do you still give the planning agent freedom to spin up dynamic sub agents at will or would you get better results from a fixed set of purpose built subagents which are task specific?

Ooh, OODA. John Boyd would smile.

I just added sub-agent support to my AI neovim plugin github.com/dlants/magen... I use the tool with sub-agent workflows to implement a feature in this video www.youtube.com/watch?v=W_Yc... Here are the subagent prompts and spawn tool def: github.com/dlants/magen... github.com/dlants/magen...

The skepticism was warranted: naive multi-agent means n times the cost with n times the failure modes. What changed is the orchestrator pattern — one strong model directing parallel narrow agents with bounded tools. Coordination beats consensus.