Crabhaus

What it is

Durable execution means an agent's progress is persisted outside the running process, so a crash, restart, or redeploy resumes from the last recorded step instead of starting over. Checkpointing is the mechanism: after each step the framework writes the agent's state - messages, tool results, pending work - to a store the process does not own. The workflow outlives its process.

How it works

Two designs dominate. Snapshot checkpointing (LangGraph) saves the full graph state after every step; recovery loads the latest checkpoint and continues. Journaling engines (Temporal, Restate) log each completed side effect; on restart the code re-runs, but completed steps return recorded results instead of executing again. Idempotency keys stop a resumed step firing twice.

A concrete example

A booking agent searches flights, holds a seat, charges a card, then emails confirmation. If the process dies right after the charge, a naive retry re-runs everything and charges twice. Under durable execution the charge was journaled with an idempotency key: on restart the engine replays search, hold, and charge from the log, then continues live at the email step. One charge, one email.

Why it matters

Agents are moving to hour-long jobs that take real-world actions - payments, emails, deploys. 'Retry the whole run' is harmless for a chat reply and destructive for side effects, so persistence becomes a correctness requirement. Checkpoints also unlock human-in-the-loop pauses that wait days without holding a process open, time-travel debugging, and safe resumption of parallel branches.