What durable execution means for agents

Durable execution means the progress of a task is saved as it happens, so that if the process dies, the task can continue from the last completed step instead of starting again.

For a short script this does not matter. For an agent that makes ten tool calls over several minutes, it matters a lot. Some of those calls cost money, some change data in other systems, and some take a long time.

Restart versus resume

Without durability, a crash means one of two things. You rerun the whole task, which repeats work and may repeat side effects. Or you give up and tell the user to try again.

With durability, the system keeps a record of what has finished. After a crash, a new process reads that record, skips the finished steps, and continues.

The event log idea

A common way to build this is an append-only log of events. Each time something completes, such as a model response or a tool result, an event is written. The current state of the run is whatever you get by replaying the events.

This gives you a few useful things:

• recovery after a crash
• a full history for debugging
• the ability to pause a run and continue it later

The cost

Replay only works if your own code behaves the same way each time it is replayed. Anything non-deterministic, such as the current time or a random number, has to be recorded as an event too. Side effects need care so that replay does not repeat them.

That is the main trade. You accept some discipline in how you write steps, and in return failures stop being disasters.