Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Why Your AI UX Is Broken (and It's Not the Model's Fault) — Mike Christensen, Ably
TL;DR
The default AI chat stack is optimized for demos, not products — Mike Christensen argues that SSE-based direct HTTP streaming in tools like Vercel AI SDK and Times back AI works for a single client talking to a single agent, but breaks once you need resilience, multi-device continuity, or real-time control.
Three capabilities separate a fragile demo from a real AI UX — The teams Ably studied across 40 companies and 10 industries kept converging on resilient delivery, continuity across tabs/devices, and live control so users can steer or interrupt agents while they work.
SSE creates a nasty resume-vs-cancel tradeoff — Because server-sent events are one-way, a client hitting “stop” can only close the connection, leaving the backend unable to tell whether it should buffer tokens for a future resume or cancel generation to stop burning money.
'Durable sessions' are the architectural shift that fixes the coupling problem — Instead of tying an agent’s output to one request and one socket, companies are inserting a persistent shared session layer between clients and agents so streams can resume, fan out to multiple surfaces, and accept input from any participant.
Multi-agent UX gets simpler when progress updates don’t flow through the orchestrator — In Christensen’s example, specialized sub-agents write directly to the shared session, so users get granular live visibility without forcing the orchestrator to also act as a progress-proxy.
Ably’s pitch is that pub/sub primitives can become AI UX infrastructure — He frames Ably channels and the new Ably AI Transport SDK as the drop-in layer that handles resumability, event materialization, multiplexing, push notifications, and even human handoff without custom plumbing.
The Breakdown
The big reveal: your model probably isn’t the problem
Mike Christensen opens with a room check: almost everyone has built some kind of AI chat app, which is exactly why his critique lands. The standard pattern — client sends a message, agent streams tokens back over SSE — is easy to ship, but he says it quietly locks you into a thin, single-client experience.
Why Ably thinks it has seen this movie at scale
He introduces Ably as a real-time messaging platform handling more than 2 billion devices, 30 billion monthly connections, and 2 trillion API operations. More importantly, he grounds the talk in conversations with engineering teams at 40-plus companies across 10 industries that are shipping AI assistants and copilots to millions of users.
The three capabilities great AI products keep investing in
Christensen says the best teams obsess over resilient delivery, continuity across surfaces, and live control. His examples are very human: walking out of the house and losing Wi‑Fi, reopening the app on your phone, or steering an agent mid-task the way people already do with Claude Code.
Where direct HTTP streaming starts to hurt
The core issue is coupling: the stream lives and dies with one client connection. If that connection drops, the stream is gone; if you open another tab or device, it can’t see the in-progress response; and because the stream is private, no other client can interrupt or guide the agent.
Resumability sounds simple until you have to build it
He walks through what it takes to resume an interrupted stream in a direct HTTP setup: store events in Redis, assign sequence numbers, build replay logic, and figure out exactly what each reconnecting client missed. His point is that your agent ends up babysitting transport complexity instead of focusing on the actual AI behavior.
The SSE trap: stopping and resuming fight each other
One of the sharpest moments in the talk is his explanation of why SSE makes “stop” ambiguous. Since it’s one-way, the client can only close the connection — so the backend has no clean signal telling it whether to keep buffering for resume or cancel the LLM and stop spending tokens; he notes Vercel’s AI SDK docs explicitly say abort is incompatible with resume.
Durable sessions: the shared layer between clients and agents
His proposed pattern is to decouple clients and agents through a persistent, stateful session that both sides connect to. That shared layer means any tab or device can stay in sync, resume automatically, and send control signals upstream while the agent is still working.
The multi-agent demo and the handoff to a human
The demo makes the architecture concrete: a support chat for an electronics shop stays synchronized across tabs, survives refreshes and even forced network disconnects, and lets one tab cancel work started in another. It gets more interesting when multiple specialized agents work concurrently in the same session, and then a human support rep joins with full context — seeing the AI conversation history and continuing the thread without the customer starting over.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
