How to Leverage Temporal for Agentic Orchestration and Workflows
summary
Instacart runs Temporal on self-hosted clusters at ~56 million workflows/day, across 45 namespaces and peaks around 200k actions/sec in a single namespace. Built systems include automated cataloging that generates table/column descriptions and opens PRs with human approval, live and twice‑daily offline eval pipelines using ensemble judges (Gemini/Claude/OpenAI), and fan‑out batch inference for embeddings and scoring. Temporal primitives—signals, child workflows, retries, heartbeats—provide the durability needed, though rate‑limiting and task‑queue tuning require careful monitoring.
AI-written summary · about
chapters (14)
- 0:00Introduction and speaker intro
- 1:11Temporal at Instacart overview
- 4:43Self-hosting Temporal architecture
- 8:56Temporal use cases at scale
- 10:19Why Temporal for AI workflows
- 13:49Temporal primitives for agentic workflows
- 15:54Three agentic workflow patterns
- 19:28Human-in-the-loop workflows
- 24:42Table metadata generation use case
- 26:49Evals for agentic workflows
- 29:15Live and offline evals example
- 32:03Batch inference workflows
- 33:26Key takeaways for agentic workflows
- 36:13Q&A on batch inference scaling
related talks
44mOrchestrate Your Business Workflows with Temporal.io
42mUsing Temporal Workflows for Seamless Dev + Research Collaboration & Continuous Algo D
50mThe end of event-based architectures? Orchestration with Temporal
26mTemporal: A Better Way to Build Modern Applications
46mThe End of Event-Based Architectures? Orchestration with Temporal
50mScaling in Space and Time with Temporal
28mDurable Code Execution With Temporal
01h 19mOpenAI + @Temporalio : Building Durable, Production Ready Agents