Agent changes keep failing the same style check. What is the best tuning?
Why B is the answer
Add the convention to custom instructions and keep the check
Teach the convention and keep the automated check. Removing the check hides the problem.
In detail
- Instructions tell the agent the convention.
- The check stays in place to enforce it.
- Never loosen a check to suit an agent.
Use case
An illustrative scenario, not a real incident.
Agent code keeps failing the style check. Adding the convention to custom instructions fixes the cause, and the check stays as protection.
See it in different ways
stateDiagram-v2 [*] --> Failed Failed --> Reasoning: right tools, wrong understanding Failed --> ToolUse: right tool, wrong arguments Failed --> Environment: missing dependency or access Reasoning --> Tune: instructions and context ToolUse --> Tune: tool definition Environment --> Tune: setup
sequenceDiagram participant P0 as Instructions participant P1 as Agent participant P2 as Style check P0->>P1: State the convention P1->>P2: Code that follows it P2-->>P1: Pass
Key point: Teach the convention, and keep the check.
Why the other options are wrong
- A. Turn the style check off for pull requests made by agents
- Turning it off removes the safeguard.
- C. Ask reviewers to fix the style by hand on every pull request
- Hand fixes never stop the pattern.
- D. Give the agent a higher limit on the number of retries it can use
- More retries do not teach the convention.