durable workflows for agents

Most agent workflow libraries focus on easy parts like model calls and retries, but the real challenge begins when workflows outlast the HTTP request that created them—spanning minutes, hours, or days while waiting for human approval, webhooks, or sub-workflows. Durable workflows, such as those provided by Runner, solve this by persisting progress, surviving crashes and restarts, and supporting safe cancellation. The key mental model is an async function with checkpoints that re-runs from the top on recovery, fast-forwarding through completed steps using persisted results.