AI Agents / Agent state and durable workflows
AI Agents: Recovery drills
Exercise crashes before dispatch, after dispatch and after commit on a disposable backend.
Environment and inputs: Durable agent workflows; LangGraph illustrates checkpoints and review interruptions.
When to use it
A restarted process must distinguish unfinished work from operations that already committed. Design state transitions together with external operation receipts, and use a disposable backend to exercise recovery before relying on it for real records.
Check the result
Compare operation counts and final state for each injected crash point.
Save and compare results
Save the test input, configuration and result. Fix the failing boundary, repeat the same test and compare it with the saved working configuration.
Scope and limits
These exercises have not been tested on your systems. Check feature support and configuration names against the installed release and linked documentation.