Durable execution is a programming model in which a program's execution state and the results of each step are recorded in persistent storage, so that when a process or server dies or the network drops, execution resumes from the point of interruption instead of starting over. A workflow run this way is called a durable workflow.
Developers write workflows as ordinary code while the execution engine handles retries, state recovery and long waits. Temporal, Cadence, Azure Durable Functions and Restate are implementations of the model. In AI it is used to run agent loops, which repeatedly call LLMs and tools, without losing progress.