A Cursor engineer explains how the company built Cloud Agents — cloud-hosted AI coding agents that each get a dedicated VM, run in parallel, and produce proof-of-work artifacts like screenshots and videos. The talk covers the evolution from a fragile homegrown runner-node architecture (essentially a bad Temporal clone) to a Temporal-powered system using Temporal Cloud, child workflows, signal-with-start for follow-up prompts, and a Kubernetes-based deployment pipeline. Key engineering details include: separating presentation state into S3/Redis to avoid expensive workflow replays, using activity markers to handle stream rewinds on retries, testing workflow upgrades in CI by replaying sampled production histories to catch non-determinism errors, and using versioned auto-upgrade with the Temporal Kubernetes controller for safe deploys. The result: activity success rate improved from ~90% to over 99%, with 50M+ Temporal actions and 7M+ unique workflows per day. One-third of internal Cursor PRs now come from Cloud Agents.