"Retry the tool call" sounds harmless until the tool issues a refund. Then durability, identity, and evidence stop being three concerns owned by three different teams and turn into one incident.
Take a support system built out of several agents. A triage agent classifies the ticket. An account agent retrieves customer state. A policy agent decides whether a refund is allowed. A human approves the recommendation. An action agent calls the refund tool. Then the process loses its connection before the response is recorded.
Retrying is the easy part. What matters is whether the platform and your own code can work out what already happened, without sending a second refund.
Restate: durable code and journaled steps
Restate is a runtime that sits in front of your services and records execution in a journal. Its AI documentation shows LLM calls and tool executions wrapped as durable steps. After a crash, completed results are replayed from the journal and execution continues from the first step that never finished. Restate supports services, virtual objects, and workflows, and runs through Restate Cloud or a server you host yourself.
