Durable Execution

The property of an agent runtime that lets a run survive crashes, restarts, and long asynchronous waits by persisting its full state and resuming exactly where it left off — potentially minutes or days later. Durable execution is what makes real Human-in-the-Loop (HITL) oversight practical: an agent can pause at an Approval Gate, persist its state, and wait for a human to respond hours later without holding a process open or losing context. The load-bearing mechanism is Checkpointing (Agent State) — snapshotting state at each step — which LangGraph provides natively and which dedicated durable-execution engines (Temporal, Restate, DBOS) provide as infrastructure. Without it, an agent that crashes mid-workflow either loses everything or re-runs side effects it already committed. Distinct from Stopping Conditions, which bound how long an agent runs; durable execution governs whether a paused or interrupted run can be faithfully resumed.