Agent Durability
Make a long agent run survive a crash: the window where a side effect lands before anything records it, idempotency keys that survive replay, and what to checkpoint.
Last updated
After this section you can
- Locate the window in every loop step where a crash duplicates a side effect, and explain why it cannot be closed
- Derive idempotency keys that survive a replay, and spot the timestamp that silently defeats them
- Persist the transcript rather than derived state, and account for the cold cache and reset budgets a resume costs
18
Agent Durability
A long agent run is a distributed transaction nobody designed as one. When it dies at step 40, the useful question is not “can I resume?” but “what did step 39 already do to the world?”
THE CENTRAL IDEA
The model is stateless: replaying a request just costs you tokens. Your tools are not. Every run therefore has a window in each step where a side effect has landed but nothing durable records that it did — and a crash inside that window is the only one that hurts. Durability is the work of making that window small, and making a replay inside it harmless.
One step of the loop — where a crash actually costs you something