Durable Execution

Agents run for minutes, hours, or days: calling tools, waiting on APIs, pausing for a human approval. In an ordinary process, a crash, deploy, or timeout in the middle loses everything, and re-running from the start can repeat side effects such as sending an email twice or charging a card twice.

Durable execution solves this by persisting progress. Each completed step and its result is saved to durable storage; if the process dies, a new worker replays the run using the saved results and continues from the first unfinished step. Workflow engines such as Temporal popularized the model, and several agent frameworks now offer checkpointing built on the same idea.

It pairs naturally with human-in-the-loop approvals, where a run may wait for days without holding a process open.

How it works

The engine logs a history of steps (calls, results, timers, incoming signals). To recover, it re-runs the workflow code from the start, but completed steps return their recorded results instead of executing again, so execution fast-forwards to where it stopped. Because replays re-run your code, workflow logic generally has to be deterministic, while side-effecting work runs in separate steps that are recorded once. Retries with backoff and timeouts are typically built in.

Example

An agent researches a topic, drafts a report, then waits for a manager's approval before emailing it. The server restarts overnight. On restart the run replays: research and draft steps return saved results instantly, and it resumes waiting for the approval. The email is sent once, not twice.

How it differs

Common misconceptions

FAQ