An agent starts editing files before anyone has seen its approach, and rework is high. What do you change?
Why D is the answer
Have it output a structured plan that is validated before it can act
Separating planning from action is a core skill here. A checked plan catches wrong approaches before they cost anything.
In detail
- The rework comes from acting before the approach is checked.
- A structured plan can be validated before the agent acts.
- Validate the plan against the task's success criteria.
Use case
An illustrative scenario, not a real incident.
An agent rewrites a service the wrong way, and a day of review is wasted. Requiring a structured plan first lets a reviewer reject the approach in minutes.
See it in different ways
flowchart LR
A["Task"] --> B["Agent writes a structured plan"]
B --> C{"Plan approved?"}
C -->|No| B
C -->|Yes| D["Agent acts"]
D --> E["Inspectable artifacts"]sequenceDiagram
participant A as Agent
participant V as Validation
participant R as Repository
A->>V: Structured plan
alt Plan passes
V->>A: Go ahead
A->>R: Makes changes in a branch
else Plan fails
V-->>A: Fix the plan first
endflowchart LR S0["Task"] --> S1["Structured plan"] --> S2["Validated"] --> S3["Agent acts"]
Key point: Check the approach before the agent acts.
Why the other options are wrong
- A. Increase its memory so it can hold more of the codebase
- More memory does not stop a wrong approach.
- B. Run it only at night so rework happens outside work hours
- Running at night moves the rework without reducing it.
- C. Give it more tools so it can check its own changes
- More tools do not check the approach.