The definitive realtime experience platform of the internet. Built for scale.
Joined June 2012
- Tweets 2,998
- Following 970
- Followers 1,864
- Likes 1,490
632 Photos and videos
Jun 2
WebSockets are the right protocol for production AI chat.
But they aren’t sufficient on their own.
And this distinction matters.
1
120
Jun 2
What the session layer provides:
-> Persistent conversation state across drops
-> Replay from last received serial (no duplicates)
-> Protocol fallback through enterprise proxies
-> Same session from any device
1
37
Jun 2
Full breakdown - why the Redis buffer approach has edges and what infrastructure actually addresses the session problem: ably.com/blog/is-websockets-…
29
May 27
The connection resumed. The conversation did not.
This is the failure mode that catches AI chat teams in production. Here's the anatomy of it.
1
156
May 27
What a session layer actually needs:
-> State stored against a stable session ID
-> Offset-based replay with no duplicates
-> Automatic protocol fallback
-> Multi-device delivery
1
52
May 27
Full breakdown with failure modes, configs, and session layer architecture: ably.com/blog/is-websockets-…
31
Ably Realtime retweeted
May 7
Ably CEO Matthew O'Riordan discusses the shift to durable sessions for long-running AI agents and solving HTTP limitations in this interview with The New Stack.
Thanks to Ably thenewstack.io/ably-durable-…
3
1
3
329
Apr 20
Generating tokens is the solved problem. What breaks in production is everything around it: streams that don't survive a page refresh, sessions tied to a single connection, agents with no signal when the user has gone.
V3 of the Ably Python SDK is built for that layer. AI Transport support is now fully available.
Check it out here: ably.com/blog/ably-python-sd…
1
150
Ably Realtime retweeted
Watching @intercom's @brian_scanlan at @aiDotEngineer London.
They set ambitious AI adoption targets and actually hit them. They're just shipping 🚀
Proud @ablyrealtime powers their realtime conversational layers. Even prouder to learn from how they build.
3
8
658
Apr 13
Introducing Ably Python SDK V3, rebuilt for production AI infrastructure.
AI Transport support is now fully available in Python: ordered token delivery, resumable sessions, multi-device continuity, and live steering. The backend work most teams end up rebuilding themselves, built in from the start.
Learn more: ably.com/blog/ably-python-sd…
2
4
144
Ably Realtime retweeted
This was so much fun.
Great talking to @cormacb and @osanseviero about open-weight models & on-device use cases. They even got Gemma running on a Nintendo Switch!
I'm excited about pairing powerful on-device intelligence with CRDTs & cross-device sync for offline-first apps.
Apr 8
As tradition, the AI Engineer speaker dinner is a great kick off before the first day of talks tomorrow!
If you're ever debated submitting a talk proposal, this dinner is "the single best networking event in London" - per @swyx
1
1
6
262
Ably Realtime retweeted
And that’s a wrap. What an amazing few days spent with the AI builders of Europe. Thanks to @swyx for bringing us all together.
The journey has only just begun.
4
3
15
3,079
Apr 9
Your AI assistant is mid-sentence explaining a complex debugging strategy. The user refreshes the page. The response restarts from the beginning, or disappears entirely. 🔄
HTTP streaming delivers tokens once, to one connection. There's no mechanism to resume from where the client left off. The application either re-invokes the LLM (paying for the same tokens twice, often getting different output) or loses everything streamed before the drop.
Users refresh pages. Networks drop. Browsers crash. Designing for the happy path means the first real-world interruption breaks the experience completely.
Read about what resumable streaming actually requires and why the naive approaches don't hold: ably.com/blog/resumable-llm-…
#LLMStreaming #AIEngineering #RealtimeAI
2
2
85
Apr 6
Orchestrators are useful… until they become the bottleneck.
In multi-agent systems, the orchestrator often ends up serialising what should be parallel. Agents finish work, but updates get stuck behind a single control plane. Users see “thinking…” while the system is already producing partial results.
The real gap isn’t more reasoning, it’s shared realtime context.
Multi-agent systems behave like distributed systems: many components updating state at once, reacting to changes, coordinating dynamically. That requires infrastructure that supports continuous updates, not just final outputs.
Find out how Ably AI Transport makes multi-agent collaboration work in realtime: ably.com/blog/multi-agent-ai…
1
49
Apr 3
Enterprises can't afford AI systems that operate behind a curtain. Every message, every agent action, every change in state needs to be visible because trust comes from knowing what's happening, not guessing.
Ably AI Transport gives you realtime observability, message-level governance, and clean integration into your existing pipelines. So you can move fast and innovate boldly, while staying firmly in control of what your AI systems are actually doing.
Discover more - ably.com/ai-transport
54