Ably AI Transport x Temporal

Temporal handles the crash. Your users still lose the thread.

Temporal keeps your agent workflows running through crashes and retries. That's durable execution. Getting workflow output to every client, whether they stayed connected or not, requires durable sessions. Ably AI Transport drops into your stack as that layer - so the AI experience you ship holds up in production.

Start building free Jump to docs

Trusted by brands serious about realtime

View customer stories

Three production gaps Temporal doesn't cover

When the workflow replays, your user gets the message again

Temporal's replay model is deterministic by design. Any send outside an activity fires again on replay. The workflow logs show one run - but the user gets the response twice.

The workflow survives the crash, but the token stream gets cut

When a worker crashes, Temporal keeps running. But the HTTP stream carrying tokens to your user's browser doesn't. That means a completed workflow still leaves the user staring at a blank screen or a half-finished reply.

The signal arrives, but the UI doesn’t know

A Temporal signal can resume a waiting workflow, but there’s no built-in way to tell the frontend it happened. The user clicks Approve, sees a spinner, and clicks again. Now there are two approvals in the workflow, and what happens next depends on whether you planned for that.

Where AI Transport comes in

Temporal handles execution. AI Transport handles everything after.

Both layers solve different problems, and together they cover what neither handles alone. Here's how the responsibilities divide and why they’re better together.

Capability	Temporal	Temporal + Ably AI Transport
Workflow durability	Temporal Yes Survives crashes and restarts	Temporal + Ably AI Transport Yes Survives crashes and restarts
Automatic retries and replay	Temporal Yes Activities retry automatically on failure	Temporal + Ably AI Transport Yes Activities retry automatically on failure
Delivering workflow output to a live client	Temporal No No built-in path to a frontend	Temporal + Ably AI Transport Yes Output is streamed as the workflow progresses
Session survives tab close / reconnect	Temporal No Client has no way to resume	Temporal + Ably AI Transport Yes Client resumes from last position on reconnect
Multi-device continuity	Temporal No Each device must poll independently	Temporal + Ably AI Transport Yes Any device joins the same session and catches up
User visibility into long-running workflow	Temporal No Users see nothing until the workflow completes	Temporal + Ably AI Transport Yes Agent state and progress stream to clients in real time
Human-in-the-loop signal (approve, redirect, cancel)	Temporal No Signals work backend-to-backend but don't reach the user	Temporal + Ably AI Transport Yes Signals travel both ways over the durable session
Multi-worker fan-in to one user session	Temporal No Each worker runs independently with no shared user session	Temporal + Ably AI Transport Yes Multiple workers publish to one session, user sees one stream
Enterprise proxy traversal for delivery	Temporal No Temporal has no transport layer for browser delivery	Temporal + Ably AI Transport Yes Automatically falls back to HTTP streaming if WebSocket is blocked
Session diagnostics for user-facing side	Temporal No Execution history only; nothing for the frontend session	Temporal + Ably AI Transport Yes Connection state, presence, and delivery events surfaced in real time
Replay-safe message delivery	Temporal No Workflow code outside activities fires again on every replay	Temporal + Ably AI Transport Yes Step IDs with AI Transport deduplicate delivery across retries and replays

“Ably gives us the reliable, low-latency AI transport we need for Messenger and Fin. No polling, no dropped messages, just a platform we can finally build next-generation AI experiences on. ”

Colin Kennedy

Principal Product Engineer

AI UX capabilities

What your Temporal workflow can do with AI Transport

Progress, as it happens

Without a session layer, the workflow runs silently. The user sees a spinner - or nothing. They don't know if it's working, stuck, or finished. AI Transport streams presence and progress events as the workflow executes, so the user sees where the agent is at every step, not just when it's done.

Reconnect without starting over

Let users drop and come back without losing anything. The session resumes from history the moment they reconnect. Without it, a dropped connection means a blank screen - the workflow keeps running, but the user has nothing to show for it.

Follow the user, not the tab

Make the result addressable from any device. A task kicked off on a laptop can complete on a phone. Without a session layer, the result sits on whichever tab the user started on - if they're still looking at it.

Approval that closes the loop

Accept approve, reject, or redirect signals from a real person and reflect them immediately in both directions - the agent resumes, the UI updates. Without it, the frontend has no way to know the signal landed.

One session, many workers

Let multiple workers and activities publish into the same user session without building fan-in logic yourself. Without it, multiple activities racing to reach the same user means ordering problems, duplicates, and logic you hand-roll and debug in production.

Send once, deliver once

Guarantee single delivery even when Temporal's replay model re-executes workflow code. Without it, a message send outside an activity fires again on replay - the workflow logs one run, the user gets it twice.

“Ably makes realtime just work. When we needed to get Breeze (our AI assistant) working in production, Ably was an obvious choice for us. It's reliable, efficient and our collaboration with the Ably team helps the engineering team build that much faster.”

Whitney Sorenson

Chief Architect, HubSpot Next

Architecture

Where Ably sits next to Temporal

Temporal owns the workflow. AI Transport owns the session.

Temporal: workflow orchestration, activities, signals, durable execution
AI Transport: durable session layer between the workflow and the client
Works with any Temporal SDK: Go, Java, TypeScript, Python, .NET
Drop-in: doesn't touch your existing workflow or activity code
Framework-agnostic on the client side: works with whatever you're using

See it in practice

How a Temporal activity becomes a durable delivery step

Each Temporal activity wraps one Step in AI Transport. The step ID is derived from the activity ID so when Temporal retries, the retry carries the same step ID and its output supersedes the failed attempt rather than duplicating it.

Jump to docs

JavaScript

import { createAgentSession } from '@ably/ai-transport';
import { stepIdFor } from '@ably/ai-transport/temporal';
async function runInferenceStep({ runId, invocationId, triggerEventId }) {
  const session = createAgentSession({ /* ... */ });
  await session.connect();
  const run = session.adoptRun({ runId, invocationId, triggerEventId });
  await run.load();
  const step = run.createStep({ stepId: stepIdFor(invocationId) });
  await step.start();
  await step.pipe(llmStream);
  await step.end();
  await session.detach();
}

Performance

Performance isn’t an aspiration. It’s engineered into every layer of our platform, from protocol optimizations to global edge acceleration.

<6.5ms

message delivery latency

30B+

connections opened monthly

Integrity

Ably guarantees data integrity, so you don’t have to reinvent resilience in your own code. Your data arrives exactly how and when it should.

5+ years

since last outage

99.999999%

message survivability

Reliability

Service interruptions cost money and trust. That’s why Ably is fault-tolerant at every level - regionally and globally. 

2B+

connected devices per month

2T+

API operations per month

Availability

Ably elastically scales with demand. Our platform is globally distributed and designed to maintain high availability under peak loads.

700+

points of presence

globally distributed regions

Start building with Ably today

Get started free