01 · Steps that don't lose state
Durable Workflows
Multi-step business logic that survives crashes, deploys, and minute-long LLM calls. Each step is independently retried and persisted. No glue code, no external queues to babysit.
tracedurable execution · 1.7s
Patterns
01Reliably run critical workflowsBreak multi-step AI pipelines and complex business logic into durable, independently retried steps.02Run long jobs on serverless hostsKeep agents and multi-step jobs running past your host's request limit, without moving off serverless.03Process large datasets without starting overUse steps, fan-out, flow control, and waits to run jobs that take hours or days, survive crashes and deploys, and never redo finished work.04Durable token streamingStream LLM tokens to the browser while running generation as a durable function, so a dropped connection or a crash never loses the response.