The missing layer
between agents & users
Agents stream tokens. Networks break. Sessions die.
There is no infrastructure layer that keeps agent-to-user sessions alive across disconnects, devices, and crashes.
The session has no home
Every boundary where a connection can break is a boundary where the session dies. Today, there is no infrastructure layer that owns these boundaries.
Disconnects
Mobile networks drop. WebSocket connections die. Users lose context mid-conversation.
Devices
Start on phone, continue on laptop. The session should follow, but it does not.
Tabs
Multiple tabs, one user, one session. Browsers make this surprisingly hard.
Agents
Agent crashes, restarts, or hands off. The user should never notice.
Time
Sessions that survive minutes, hours, days. Not just the current connection.
Modalities
Text, voice, video, tool calls. One session spanning every interaction type.
Industry consensus
“The incumbent approach has been considered inadequate”
“The current transport mechanisms available for AI chat applications are inadequate for production use cases.”
Vercel ChatTransport proposal
“A pluggable transport layer is needed to abstract away the complexity of maintaining persistent connections.”
TanStack ConnectionAdapter RFC
“Real-time streaming and session management remain the most requested features from production users.”
CrewAI Issue #3259
What is a Durable Session?
A session that survives everything the network, the device, and the infrastructure can throw at it.
Think of a phone call that never drops. You walk through a tunnel, switch from WiFi to cellular, hand your phone to someone else. The call continues. That is a durable session, but for agent-to-user communication.
Reliable Streaming
Ordered, resumable, exactly-once delivery across any transport. Never lose a token.
stream.resume(lastEventId)Session Continuity
Survive disconnects, device switches, tab changes. The session outlives the socket.
session.attach(newDevice)Agent Visibility
Agents see connection state. They know when users disconnect and when they return.
agent.onPresenceChange(cb)Convergence
Across frameworks, infrastructure providers, and AI platforms, the same pattern is emerging. Everyone is building toward durable sessions.
References
“Local-first sync provides the durability layer that AI applications need for reliable real-time streaming.”
James ArthurElectricSQL
“Durable execution ensures workflows survive infrastructure failures, but the client transport layer remains unsolved.”
Temporal$5B valuation
“We need durable execution for the backend and durable sessions for the frontend. They are complementary layers.”
Stephan EwenRestate
“Serverless infrastructure must handle persistent connections without persistent servers. This is the session durability problem.”
UpstashServerless
“Streaming AI responses over HTTP is fragile. The industry needs a transport layer that handles reconnection natively.”
Vercel AI SDKFramework
37-vendor analysis
lack reliable streaming recovery
have no session continuity
offer no agent-side visibility
Not a feature. A layer.
Durable Execution and Durable Sessions solve different problems at different layers of the stack. Both are necessary.
Durable Execution
- Backend workflow orchestration
- Survives server crashes
- Replays deterministic functions
- Temporal, Restate, Inngest
Durable Sessions
- Client-to-agent transport
- Survives network failures
- Resumes streams, maintains presence
- Ably, ElectricSQL (emerging)
Developer voices
The same pain, reported independently across the ecosystem.
“We need a way to resume streams after disconnection without replaying the entire conversation.”
Mastra
“The transport layer is the most underinvested part of the AI stack. Everyone rebuilds it from scratch.”
Lars Grammel, Vercel
“Connection state management is eating our development time. This should be infrastructure.”
Pydantic AI
“Real-time bidirectional communication with AI agents requires infrastructure that does not exist yet.”
OpenAI Realtime
“Session persistence across worker restarts is a fundamental requirement we keep reimplementing.”
Temporal
“Our users lose context every time their connection drops. We need session-level durability.”
CrewAI #3259
“Voice AI sessions need sub-100ms reconnection. HTTP polling is not an option.”
Retell AI
“We built our own reconnection logic. It took months. It still has edge cases.”
Supabase
Try it out
Try one of the current durable session providers.
AI Transport - durable sessions as managed infrastructure
ably.com/ai-transport
Durable Streams protocol and sync engine
electric-sql.com
The category is emerging
Get involved
Durable Sessions is an emerging category. Help shape the definition, contribute research, or reach out to discuss.