Over the past year, I've spoken to more than 40 engineering teams building production AI agents. Different companies, different frameworks, different use cases. The same conversation kept happening.
"Our streams break when users switch tabs." "We can't tell if the agent crashed or is still thinking." "We built a custom reconnection layer and it took three months." "Our users can't switch from laptop to phone mid-conversation." Every team described it differently, but they were all describing the same gap. Between the agent and the user, there's no dedicated infrastructure for the session itself.
We backed this up with research across 37 AI infrastructure platforms, hundreds of GitHub issues and community threads, and 40+ customer discovery calls. 35 of those 37 platforms have no stream resumption after a disconnect. 33 have no way to detect an agent crash. The gap is universal, and the framework maintainers know it. Vercel built a pluggable ChatTransport in AI SDK 5 so developers can bring their own transport. TanStack AI shipped a ConnectionAdapter for third-party providers. They've diagnosed the problem and built the plugin points. They're waiting for specialist infrastructure to show up.
Nobody did anything wrong. Everyone focused on the right thing first: the intelligence, the orchestration, the models. But as AI experiences have gotten more sophisticated, the transport layer between the agent and the user has become the constraint.
Agents are becoming human-like, and they need human infrastructure
The insight that changed how we think about this came from an unexpected direction. As agents get more sophisticated, they start behaving like human participants in a conversation. They think for a while before responding. They work on tasks in the background. They hand off to a human colleague when they hit their limits. Users walk away, come back later, and expect to pick up exactly where things were.
These are the exact communication challenges we've been solving for human-to-human interaction for 10 years. Presence, reliable delivery, session continuity across devices, bidirectional control. Every messaging app since WhatsApp has solved these problems for humans, and the moment agents become participants in conversations, they need the same infrastructure.
We've been building this for a decade
I'll be honest. We almost dismissed the AI space entirely. When every company suddenly needed an "AI strategy," my instinct was skepticism. We're an infrastructure company. We process trillions of transactions across billions of devices. Why would we need an AI-specific product?
I was wrong. Companies like Intercom and HubSpot were already building AI agent experiences on top of our Pub/Sub messaging infrastructure, the realtime layer that handles reliable delivery between servers, devices, and services. They needed ordered delivery, presence, session state, multi-device support. They were using the infrastructure we'd already built, without waiting for us to package it as an AI product.
Ably has been a durable session layer for 10 years. We never called it that because the term didn't exist. We called it realtime infrastructure, messaging, pub/sub. But the capabilities are the same. Persistent sessions that survive disconnects. Ordered delivery with automatic catch-up. Multi-device fan-out, presence, bidirectional communication. We built all of this for human communication at scale, and it turns out it's exactly what AI-to-human communication needs too.
A category is forming
We're not inventing this term. We're recognizing something that's already happening.
ElectricSQL published a "Durable Sessions" blog post earlier this year defining it as a pattern for collaborative AI. EMQX has used "Durable Sessions" as a named feature in their MQTT broker for years. Convex is building agent components with persistent threads and durable workflows. Vercel is building a DurableAgent class. At least 12 companies are converging on the same problem space from different angles.
The pattern mirrors Durable Execution. Temporal existed before AI agents needed it, then suddenly every team building production agents needed backend workflows that couldn't fail. Temporal went from niche to a $5 billion valuation. AWS adopted the term for Lambda Durable Functions. The category debate was over.
Durable Execution made the backend crash-proof. Durable Sessions makes the experience crash-proof. They're complementary layers on opposite sides of the agent.
This needs to be bigger than Ably
A category with one company in it isn't a category. It's a product pitch. We want other companies in this space. We want developers to recognize "durable sessions" as an infrastructure layer they need, regardless of who provides it.
We've published durablesessions.ai as a community resource that defines the concept, documents vendor convergence, and tracks how the ecosystem is forming. I'm personally committed to pushing this forward. Not because it helps Ably specifically, but because I believe it will improve how we all build and experience AI. I've been doing this for a long time and I've never been more energized about what's ahead.
If you're at AI Engineer Europe next week, our tech lead will be presenting on durable sessions and why this layer matters. I'll be there too. Come find me and the team. If you're building in this space, whether as a competitor, a complement, or a fellow traveler, I want to talk. Getting the people working on this in the same room, having honest conversations about what developers actually need, is worth more than any blog post.
This is the first in a series. Over the coming weeks, we'll go deeper into the evidence, the ecosystem, and the practical framework for evaluating what your AI sessions actually need. Follow along here or connect with me on LinkedIn.
