Building Agentic Consumer Products with Temporal
summary
Cursor builds cloud agents that run each job in a dedicated VM per agent and orchestrates them with Temporal Cloud workflows, using worker activities on AWS (migrated from ECS to Kubernetes/EKS), S3 for content-addressable artifacts, Redis streams for live output, VNC for proofs, and child workflows for sub-agents. Reliability jumps from ~90% to >99% activity success, at ~50M Temporal actions and 7M+ workflows/day, one third of internal PRs coming from agents, and CI replays production histories to catch workflow upgrade nondeterminism.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:09AI coding industry eras
- 4:17Cloud agents demo
- 5:22How local agents work
- 7:11How cloud agents work
- 8:33Why cloud agents level up
- 9:35Temporal for agent reliability
- 10:37Road to cloud agents
- 13:23Current architecture with Temporal
- 16:10Sub-agents and child workflows
- 17:12Presentation layer design
- 19:40Temporal infrastructure evolution
- 24:59Testing workflow upgrades
- 26:22Cloud agents by the numbers
- 27:45Other Temporal use cases
- 31:18Conclusion and Q&A
related talks
33mFrom Prototype to Production: Making Agentic AI Systems Reliable
44mBuilding Agentic Workflows in Cursor IDE
24mDurable Agents: long running AI workflows in a flakey world
01h 24mBuilding your own software factory
42mUsing Temporal Workflows for Seamless Dev + Research Collaboration & Continuous Algo D
32mHow I Build Web Applications in 2026
52mDesigning Production-Ready Agentic Systems: Lessons from the Field
25mEngineering Production-Ready Onboarding Agents