So you want to build AI Agent group chat like OpenAI? (zknill.io)

On Nov. 13 OpenAI piloted group chats in ChatGPT, prompting a look at why mainstream agent frameworks (Vercel AI SDK, Google ADK, LangChain, Mastra, Anthropic, OpenAI SDKs) struggle to replicate that feature. Most systems use long-lived HTTP connections with Server-Sent Events (POST-and-wait streaming): clients open a connection, the model streams token deltas, and the client reassembles the response. That pattern breaks down for multi‑participant chats—if a client disconnects, the streamed work is lost; stateless backends face load‑balancer routing, context provisioning, and database‑heavy polling problems; and fan‑out to multiple subscribers becomes clumsy and inefficient. Realtime Pub/Sub architectures address these gaps. Treat each group chat as a channel where participants (humans and agents) publish and subscribe: channels handle message fan‑out, persistence/history for late joiners and model context, presence to track active users, and shared state (e.g., “LiveObjects”) for collaborative tasks. That means agents can access full conversation history without re-sending context each request, servers avoid brittle long‑lived HTTP ties, and clients get true realtime updates instead of polling. For teams building agentic group chat, Pub/Sub offers a practical, scalable alternative to the ubiquitous SSE streaming pattern.