Inside HeyGen’s durable agent harness, powered by Temporal

HeyGen replaced a custom video-generation orchestration system built on MySQL, RabbitMQ, Celery, polling, and callbacks with Temporal to improve reliability, recovery, and development speed across a complex pipeline involving speech generation, GPU avatar inference, rendering, compositing, billing, and delivery. Temporal workflows durably coordinate sequential and parallel activities, retries, compensation such as quota refunds, and recovery from worker failures, while concurrent scene processing uses stage-specific limits and selectively applies child workflows for isolated timeout or failure boundaries. At scale, the platform manages millions of workflow executions and tens of millions of activities daily through roughly 300 workflow types, nearly 1,000 activities, about 100 task queues, and more than 50 independently scaled worker deployments across multiple GPU providers. HeyGen also built shared tooling for identity propagation, retry policies, metrics, visibility, heartbeats, worker setup, and deterministic workflow testing, alongside admission controls and task priorities that enforce customer fairness. To control GPU costs and capacity, it uses workload-specific queues, a Temporal-based global autoscaling coordinator that accounts for worker readiness, demand, provider cost, and allocation priorities, plus a custom slot mechanism that overlaps input preparation with inference without allowing concurrent GPU use.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.