An agent's output fails review partway through a chain. What should happen to later steps?
Why A is the answer
They stop until the output is corrected, or the step is rolled back
Gates between steps keep bad output from spreading. Roll back to a known good point when needed.
In detail
- Later steps depend on this output.
- Stop them until it is corrected, or roll the step back.
- Roll back if the correction would take too long.
Use case
An illustrative scenario, not a real incident.
A design step fails review. Later steps wait until it is corrected, so no work is built on a known fault.
See it in different ways
stateDiagram-v2 [*] --> Running Running --> Stalled: no progress Stalled --> Running: retry from a checkpoint Stalled --> Escalated: retry fails Escalated --> [*]
sequenceDiagram participant W as Workflow participant A as Agent participant H as Human W->>A: Step A--xW: Stalls or fails review W->>W: Pause later steps W->>A: Retry or roll back W->>H: Escalate if it still fails
stateDiagram-v2 state "Step output" as s0 state "Review" as s1 state "Later steps" as s2 state "Correct or roll back" as s3 [*] --> s0 s0 --> s1 s1 --> s2: passes s1 --> s3: fails s3 --> s0 s2 --> [*]
Key point: Stop later steps until the fault is fixed.
Why the other options are wrong
- B. They restart from the very beginning of the whole workflow
- Restarting everything throws away good work.
- C. They continue, and any problems are repaired after they finish
- Later steps would build on a known fault.
- D. They run on the failed output, marked with a warning for later
- A warning does not make the output usable.