An emerging infrastructure category

The missing layer
between AI and humans

Durable Execution made backends crash-proof.
Durable Sessions makes the experience crash-proof.

Understand the gap Who's talking about it
Agent A LLM + Tools Agent B Orchestrator Agent C Specialist Durable Sessions Streaming State Presence Lifecycle Resilience Laptop connected Phone reconnecting... Tablet connected AI AGENTS HUMAN DEVICES session persists
The problem

The session has no home

AI agents have state. Durable execution platforms persist workflow state. Databases store data. But it's not clear which layer in the AI infrastructure stack should own the session — the stateful interaction between an agent and the humans it serves.

Disconnects

Stream dies on network drop. User restarts from scratch.

Devices

Conversation started on phone can't continue on laptop.

Tabs

Opening a second tab creates a second, disconnected session.

Agents

Multi-agent handoff loses context. Crashed agent looks like a thinking one.

Time

Background task completes after user leaves. Result is lost.

Modalities

Voice call can't transition to text without losing state.

Every team building serious AI products hand-rolls fragile session infrastructure from scratch.

The signal

The incumbent protocol has been declared inadequate

The MCP protocol deprecated SSE in favour of Streamable HTTP. Vercel built a pluggable ChatTransport interface and told the community to bring their own WebSocket implementations. TanStack AI shipped a ConnectionAdapter designed for third-party transport providers. AG-UI built pluggable transport from the ground up.

The ecosystem has diagnosed the problem, designed the plugin points, and is waiting for someone to show up with the infrastructure.

Definition

What is a Durable Session?

A durable session is a persistent, addressable session that sits between agents and users as the medium through which they interact. It's not a connection — connections break. It's not a channel — channels are a transport primitive.

A durable session is the stateful layer that outlives any single connection, and that any participant — user, agent, device, service — can join, leave, and return to over time.

The analogy
Durable Execution made your backend crash-proof
Durable Sessions makes the experience crash-proof

Where it sits in the stack

Agent Layer LLMs · Tools · Memory · Durable Execution
Durable Sessions Streaming · State · Presence · Lifecycle
Client Layer UI · Devices · Tabs · Modalities
Convergence

Multiple vendors, same conclusion

Independently, across different stacks and use cases, companies are converging on the same missing layer. Nobody orchestrated this. The pattern emerged.

ElectricSQL
Published a "Durable Sessions" blog post with reference implementation using Durable Streams + TanStack DB
January 2026
EMQX
"Durable Sessions" as a named MQTT feature — session state surviving broker restarts
Production feature
Vercel
Building @ai-sdk/durable-agent. Pluggable ChatTransport in AI SDK v5
In development
Ably
AI Transport — durable sessions as the infrastructure layer between agents and human devices
March 2026
Upstash
"Resumable AI SDK Streams" using Redis Streams for stream recovery
Published solution
Convex
Agent component with persistent threads, durable workflows, and real-time sync
Agent component
AG-UI
Pluggable transport architecture. The protocol moved beyond fixed transport assumptions.
2025
TanStack AI
ConnectionAdapter interface — explicitly designed for third-party transport providers
Documentation
MCP Protocol
Deprecated HTTP+SSE transport in favour of Streamable HTTP. The protocol layer itself evolving.
May 2025
35/35
AI infrastructure vendors with no stream resumption at the infrastructure level
33/35
vendors with no agent health signal or crash detection
32/35
vendors with no multi-device fan-out
26/35
vendors where SSE fragility is a documented developer pain point

Based on a cross-vendor synthesis of 35 AI infrastructure vendors, March 2026

Vendors analysed

Ably PubNub Pusher EMQX ElectricSQL Vercel AI SDK LangChain LangGraph TanStack AI AG-UI Temporal Restate Inngest Upstash Convex Supabase Firebase CrewAI Cloudflare Agents LiveKit Stream Sendbird Twilio Socket.IO Centrifugo NATS RabbitMQ Apache Kafka Azure SignalR AWS API Gateway Google Pub/Sub CometChat Knock Courier MCP Protocol
An emerging AI infrastructure layer

Not a feature. A layer.

A feature is "sessions that don't break." An infrastructure layer is what makes it possible to build AI experiences that work across every boundary — disconnects, devices, agents, modalities, and time.

Durable Execution e.g. Temporal

  • Decouples application logic from infrastructure failure
  • What outlives failure: the execution
  • What stays stateless: your backend code
  • Mental model: "Write code as if failure doesn't exist"

Durable Sessions

  • Decouples agent interaction from transport failure
  • What outlives failure: the session
  • What stays stateless: your agent and your client
  • Mental model: "Build AI experiences as if disconnects don't exist"
References

Further reading

The conversation is happening across blog posts, GitHub issues, protocol specs, and documentation.

ElectricSQL Durable Sessions Blog post · Jan 2026 EMQX Durable Sessions — MQTT Session Durability Documentation Ably AI Transport — Durable Sessions Infrastructure Product · March 2026 Vercel AI SDK ChatTransport & Durable Agent Discussion GitHub Issue #6502 TanStack AI ConnectionAdapter — Pluggable Transport Documentation Upstash Resumable AI SDK v5 Streams with Upstash Realtime Blog post Convex Agent Component — Persistent Threads Component AG-UI Protocol Architecture — Transport Agnostic Design Protocol documentation