You’ve invested in prompt engineering, evals, fine-tuning. But the orchestration layer gets almost no attention. What happens when your agent is mid-run and an API goes down? When a long-running agent loses state? When retries cause duplicate actions? These are distributed systems problems, not model problems. We built Output.ai, an open-source TypeScript framework with Temporal at its core, to make durable agent execution accessible to any TypeScript developer. We extracted it from production: agents powering growth workflows at Lovable, Webflow, Airbyte, and dozens of other teams that needed agents to recover, replay, and run unsupervised.