Durable Execution

Durable execution means the runtime records a job’s progress at each point it waits, and nowhere else. A wait is a tool call, an external event, or a human decision. If the process dies between waits, it resumes from the last recorded point instead of starting over. Nothing before that point runs twice.

Durability is decomposition. You split the job at its waits, and each segment between two waits is a unit that either finished or did not. A retry replays the job and skips every segment that already finished, which only works if each step is idempotent, safe to run again with the same result.

Durable execution is also what makes real human in the loop possible. A local session executes once, start to finish, while you watch; it cannot pause for three days waiting on an approval. A durable runtime checkpoints at that wait and resumes when the answer arrives, in three seconds or next week.

The durable runtime is the host a local session cannot be: the supervisor that holds the retries, the waits, and the logs a person otherwise has to sit and hold. Pausing across real time and resuming is the reason to move an automation onto it; the crash recovery comes with it.