An agent resumes after a pause and repeats finished steps. What prevents this?
Why D is the answer
A progress artifact that records what is done and what remains
Persisted state lets an agent resume without repeating or diverging from earlier work.
In detail
- A progress artifact records what is done and what remains.
- The agent reads it when it resumes.
- Update it after each step, not only at the end.
Use case
An illustrative scenario, not a real incident.
An agent stops overnight and repeats two steps next morning. A progress file listing what is done and what remains prevents that.
See it in different ways
flowchart LR
A["Long-running work"] --> B["Save progress to an artifact"]
B --> C{"Still matches the plan?"}
C -->|Yes| A
C -->|No| D["Correct, or reset from the artifact"]mindmap
root((Agent memory))
Short term
This task only
External
Files in the repository
Progress artifacts
Copilot Memory
Repository facts
Care
Expiry
Pruningmindmap
root(("Progress"))
n1["Done"]
n2["Finished steps"]
n3["Remaining"]
n4["Next steps"]
n5["Decisions"]
n6["Choices made"]Key point: Record what is done and what remains.
Why the other options are wrong
- A. A larger context window on the next run of the same agent
- A larger window does not record what was finished.
- B. A higher temperature setting, so that it picks a different path
- More randomness may repeat or skip steps.
- C. A shorter task description that leaves out the finished steps
- A shorter description does not record progress.