Your @OpenAI agent's answer now survives a refresh or a dropped connection, because the reply lives on a session instead of on the connection. Any device can pick the same thread back up. Demo: ably.com/blog/ai-transp…#AIAgents
Guaranteeing WebSocket order in production isn't a one-time fix. TCP's guarantee ends at one connection. So past that, sticky routing, a shared coordination layer, and replay sequencing are system requirements. But should you manage the infra?
Fin's Colin Kennedy, on deciding whether to keep maintaining your own WebSocket infra for AI streaming: "Is maintaining this system helping you build a better product?"
The framework to work through that decision:
Building LLM streaming? You've probably got session IDs, a Redis buffer, and dedup logic. That's five systems you maintain forever, and there's a sneakier failure too: streams that look fine while nothing confirms delivery. The fix: ably.com/blog/resume-to…#AIEngineering
Your AI agent goes quiet mid-task. Thinking, stalled, or dead? From the client, you can't tell.
That's because most AI infrastructure has no signal for it, as Vercel's own maintainers have said.
Ably AI Transport gives you one.