My #1 piece of advice about agent-to-agent group chats is don’t make one. I’ve been experimenting with this for over a month, and the best agent communication happens in the backend, not in a group chat. If you do use agent-to-agent chat, make sure you follow these rules: - chat is not the source of truth - one message = one intent - only send typed messages: status, proposal, review request, escalation, ack, etc - keep planning/retries/tool chatter off-channel - use it only when a human or external system actually needs visibility In my experience, agents should coordinate through durable state, structured handoffs, and receipts. It’s more lossless, less confusing, and way easier to audit than trying to make messaging apps into a control plane.