A team assigns Copilot coding agent an issue to improve an API endpoint. The issue says only “make the endpoint faster,” and Copilot opens a draft pull request. The PR changes database queries, caching behaviour, and an unrelated authentication middleware file. The checks pass, but reviewers disagree about whether the task is complete or in scope. Which change should the team make before assigning similar tasks again?

Defining measurable outcomes and file-scope constraints is crucial because it provides clear boundaries and success metrics for the AI agent. The original prompt "make the endpoint faster" is ambiguous and led Copilot to make broad, potentially out-of-scope changes. Measurable outcomes (e.g., "reduce average response time by 20%") and file-scope constraints (e.g., "only modify apiservice.py and databasequeries.py") would guide Copilot to focused, verifiable improvements, preventing unnecessary changes to unrelated components like authentication middleware. Requiring only one commit is an arbitrary restriction that doesn't address the core problem of unclear task definition. Using passing checks as acceptance criteria is insufficient; checks only confirm code compiles and basic tests pass, not that the task's intent or scope was met. Disabling draft pull requests doesn't solve the underlying issue of poorly defined tasks; it merely changes the PR workflow.