Durable execution breaks a workflow into recorded steps. After each important transition, the system preserves enough state and event history to know what completed, what is waiting, and what should happen next.
This matters for long-running AI work because tools, services, and machines can fail while a task is active. A durable workflow can retry a failed step or resume from a known point without repeating successful work or losing the surrounding process.
ELI5
Durable workflow execution saves the state of a long-running process so it can survive delays, crashes, and restarts. The work resumes from a known point instead of starting over or forgetting what already happened.
For example, an AI research workflow can save completed searches and drafts before waiting for a person's review. If a worker restarts overnight, it can continue from the review step without repeating every earlier action.
What does a durable workflow record?
It can record completed steps, current state, pending timers, external requests, decisions, and the next action to run.
Does durable execution prevent every failure?
No. It makes failures recoverable by preserving progress and defining how retries and resumptions should work.
