An evaluation suite shows that a custom Copilot cloud agent frequently opens draft pull requests that pass linting but fail accessibility review. The failures cluster around React form components in web/forms/**. The team wants a defensible tuning loop rather than a one-off prompt change. Which four actions should you perform in order? 1. Add path-specific instructions for web/forms/**. 2. Measure failures by category and affected path. 3. Rerun the same evaluation cases. 4. Mark all existing draft PRs ready for review. 5. Add or update PR validation for accessibility checks. 6. Remove branch protection until the agent improves.

The correct sequence addresses the problem systematically. First, "Measure failures by category and affected path" (2) is crucial for understanding the problem's scope and specific areas of concern, providing data-driven insight. Next, "Add path-specific instructions for web/forms/" (1) directly targets the identified problem area with tailored guidance for the agent. Then, "Add or update PR validation for accessibility checks" (5) establishes an automated gate to prevent future regressions and enforce the desired quality standard. Finally, "Rerun the same evaluation cases" (3) is essential to verify if the changes have effectively resolved the issue and improved the agent's performance. Option 1 is incorrect because it skips the initial measurement, leading to potentially untargeted solutions. Option 2 is incorrect because marking all existing PRs ready for review (4) is premature and risky without addressing the underlying issues. Option 4 is incorrect because removing branch protection (6) would lower quality standards and is counterproductive to improving agent performance.