Checkpointing and Recovery in Multi-Step Agents

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-agentic-agent-checkpointing-recovery-creators

1. Why are idempotency keys essential when implementing checkpointing for financial transactions?

   1. They automatically retry failed transactions
   2. They ensure the transaction completes faster
   3. They prevent duplicate charges if a step partially succeeds before a crash
   4. They encrypt the transaction data

2. What does it mean for an operation to be idempotent?

   1. Running the operation multiple times produces the same result as running it once
   2. The operation can be undone after execution
   3. The operation always completes successfully
   4. The operation runs faster when repeated

3. What is the primary limitation of checkpointing that AI cannot overcome?

   1. AI cannot make every tool call automatically idempotent
   2. Checkpointing cannot be used with external APIs
   3. AI cannot detect when a crash has occurred
   4. Checkpointing slows down agent execution significantly

4. What information should be included in a checkpoint to enable proper recovery?

   1. Only the user's original request
   2. Only the final output of each step
   3. The run ID, step number, tool call details, and partial results
   4. Just the timestamp of each execution

5. What does the lesson mean by 'durable' state?

   1. State that is encrypted by default
   2. State that exists only during the current session
   3. State that persists across crashes and restarts
   4. State that cannot be modified by the agent

6. What is the purpose of defining a recovery action for each step that mutates external state?

   1. To automatically delete failed steps
   2. To make the agent run faster
   3. To specify what to do if the step must be retried after a crash
   4. To permanently store the step's results

7. What is a 'stable run ID' and why is it important?

   1. A timestamp used to log when tools are called
   2. A random number generated for each tool call
   3. A unique key for each user session
   4. A consistent identifier for an entire agent execution that persists across restarts

8. What is the risk of a 'naive resume' after a tool call partially succeeds but the agent crashes?

   1. The checkpoint will be automatically deleted
   2. The agent will ask for user confirmation to continue
   3. The tool call will be retried, potentially causing duplicate external effects
   4. The agent will skip that step entirely

9. Why should checkpoint history be surfaced to operators?

   1. To provide visibility into what happened before a crash and support debugging
   2. To automatically generate reports for compliance
   3. So operators can manually approve each step
   4. To reduce the cost of running agents

10. What external state changes must be tracked for proper checkpoint design?

    1. Only changes to the agent's internal variables
    2. Only changes that cost money
    3. Every step that modifies data outside the agent (databases, APIs, files)
    4. Only changes that are reversible

11. What does 'resumability' refer to in agentic workflows?

    1. The ability to speed up long-running tasks
    2. The ability to continue execution from the last successful step after an interruption
    3. The ability to pause an agent at any time
    4. The ability to run multiple agents in parallel

12. When implementing checkpointing for an agent that sends emails, what must you ensure?

    1. The agent must send a second confirmation email after each send
    2. The email-sending step must be idempotent or use an idempotency key to prevent duplicate sends
    3. Checkpointing will automatically delete failed emails
    4. Emails should never be checkpointed because they are fast

13. What distinguishes checkpointing from simple logging?

    1. Logging records what happened; checkpointing enables resuming execution from that point
    2. Checkpointing is faster than logging
    3. Checkpointing and logging are the same thing
    4. Logging is only for errors; checkpointing is for all steps

14. Why is it insufficient to just save the 'result' of each step in a checkpoint?

    1. You also need to know whether the step actually completed or crashed mid-execution
    2. Results are automatically deleted after 24 hours
    3. Results are always too large to store
    4. Saving results violates data privacy regulations

15. What happens if you do not define idempotency keys for a payment-processing step in your agent workflow?

    1. If the agent crashes after the payment processes but before recording success, resuming may charge the customer twice
    2. Payments will be processed faster without idempotency keys
    3. The payment will fail automatically
    4. The agent will refuse to process payments