Anthropic tested different ways of sub-agent delegation 🙂 https://x.com/jyangballin/status/2064420216025891295/photo/2

I wonder how often it will be worth it to do multi-agent:

the score improvement and latency gain come from spending more tokens and working on the problem concurrently

I'd be interested to know how/if they modeled a shared agent communication channel. I've seen this help in practice, but it opens some more complexity doors. The geometry of how information flows through an agent swarm greatly impacts the outcome.

I've been experimenting with an AgentMail MCP, an async communication channel with no guarantees, but also no much requirements. My agents were using it to ask e.g. the implementor asked the reviewer for review; the triager got upset at everyone for missing AI reviews on GitHub so sent them a note

I ran some quick experiments for my course and had trouble finding a good real world scenario for sub agents where you save money /tokens. There are some natural fits like code review, but I need better tasks to bring this out.