From prototype to production: building reliable long-running agents with Temporal + Arize
AI agents don't behave like ordinary request-response applications. They can run for minutes or hours, call multiple models and tools, and carry state across long-running workflows. A crash, deploy, or failed step shouldn't mean starting over.
Temporal makes those workflows durable by persisting each step, automatically retrying failures, and resuming interrupted sessions. But keeping an agent running is only half the problem. You also need to understand what it did, why it did it, and whether a change actually made it better.
This live session shows how Temporal and Arize work together to build reliable AI agents in production. We'll walk through the WithLove Gift Shop, an AI shopping agent running as a durable Temporal workflow and instrumented with OpenTelemetry. We'll trace the full agent run in Arize Phoenix and Arize AX, inspect its workflow steps, model calls, and tool calls, then use those traces to run an experiment and evaluate whether a change improved the agent.
What we'll cover
Durable Temporal workflows for long-running agents
OpenTelemetry instrumentation on a live shopping agent
Tracing the full run in Phoenix and Arize AX
Inspecting workflow, model, and tool calls
Running an experiment from production traces to measure whether a change improved the agent
Who this is for AI engineers and agent builders shipping agents into production, especially anyone building durable, long-running agents on Temporal (or curious what durable execution buys them), plus PMs and technical leads responsible for agent quality and reliability. Especially relevant if you want observability that works the same way in local development and in production.
You'll learn how to
Build durable, resumable agent workflows with Temporal
Instrument agents with OpenTelemetry
Trace agent behavior across Phoenix and Arize AX
Debug workflow, model, and tool behavior end to end
Turn production traces into experiments that measure whether your agent actually improved
Format
Live session with demos across Temporal, Phoenix, and Arize, plus Q&A (about 45 minutes: roughly 30 minutes of presentation and live demo, 15 minutes of Q&A).
Level Intermediate. Familiarity with LLM agents, traces, and basic evaluation concepts helps. No prior Arize, Phoenix, or Temporal experience required.