A plan fails at step three because a tool result was unexpected. What should a good agent do?
Why B is the answer
Revise the plan based on the new information
Replanning on new observations is the point of a planning loop. Unbounded retries are a known failure mode.
In detail
- A good agent revises the plan based on what it learned.
- Limit how many times it can replan.
Use case
An illustrative scenario, not a real incident.
A flight search returns nothing, so an agent tries nearby airports, rather than repeating the same call.
See it in different ways
stateDiagram-v2 [*] --> Reason Reason --> Act Act --> Observe Observe --> Reason: not done Observe --> [*]: done
mindmap
root((Memory))
Short term
Current context
Summaries
Long term
Episodic
Semantic
Planning
Plan and execute
Revise on new factsstateDiagram-v2 state "Plan" as s0 state "Act" as s1 state "Observe" as s2 state "Replan" as s3 [*] --> s0 s0 --> s1 s1 --> s2 s2 --> s3: unexpected result s3 --> s1
Key point: Replan on new information.
Why the other options are wrong
- A. Repeat step three until the tool returns the expected result
- Repeating expects a different result.
- C. Finish the remaining steps anyway and flag the problem at the end
- Later steps may depend on this one.
- D. Stop and delete all progress so the next run starts clean
- Deleting progress wastes work.