How durable agent runs actually work

A durable run is one that can survive a crash, a redeploy, or a flaky dependency without losing state or repeating side effects. The hard part is that last clause.

Checkpoint after intent, act once

Nebula records a step's intent before it executes, then marks it complete after. On resume, a step that was recorded-but-not-completed is retried; a completed one is skipped. Tools declare their idempotency so we know which are safe to replay.

The human in the middle

When a guardrail pauses a run, the checkpoint is the pause. The run is suspended to storage and resumes the moment a decision arrives, cleanly, with no polling loop and no lost context.