Descriptions:
Mike Christensen, staff engineer at Ably, presents a detailed technical breakdown of why the default HTTP streaming model used in most AI chat applications fails in production — and what engineering teams are doing to fix it. Drawing on conversations with more than 40 companies across 10 industries, Christensen identifies three foundational capabilities that distinguish polished AI products from brittle demos: resilient delivery across dropped connections, session continuity across devices and tabs, and live bidirectional control over agents while they are running.
The root problem, he explains, is that frameworks like Vercel’s AI SDK and LangChain’s AI SDK default to Server-Sent Events (SSE), which create a point-to-point connection tying stream health directly to a single client connection. When a mobile user switches networks, refreshes the page, or opens a second tab, the stream is lost. Christensen also surfaces a specific, documented incompatibility in SSE-based implementations: because SSE is strictly one-way, stop/cancel functionality is fundamentally mutually exclusive with stream resumability — a limitation explicitly noted in Vercel’s own SDK documentation.
The proposed architecture decouples agent execution from client connection state using durable sessions backed by a real-time messaging layer. This enables clients to reconnect and replay events mid-stream, and opens the door to bidirectional WebSocket transport for richer interactions — including steering or canceling an agent from multiple devices simultaneously. Ably’s platform handles over 2 billion devices and 2 trillion API operations monthly, grounding the talk’s recommendations in real production scale.
📺 Source: AI Engineer · Published May 17, 2026
🏷️ Format: Deep Dive
