Question 17
Domain 2A coordinator is asked to orchestrate a large refactor touching dozens of files. An architect wants to catch decomposition mistakes — missed goals, duplicate sub-tasks, infeasible tool assignments — before any executor starts touching source code. Which validation technique most directly addresses this, short of writing a full human-in-the-loop review process?
Correct answer: A
Explanation
Plan mode exposes the proposed decomposition before execution, so reviewers can catch “missed goals, duplicate sub-tasks, infeasible tool assignments” before any source code changes begin. This matches the validation goal of checking the plan itself first, rather than relying on a full human-in-the-loop review after work has started.
Why each option is right or wrong
A. Use plan mode to surface the proposed decomposition for approval before any tool executes.
Plan mode is the pre-execution validation step: it presents the proposed task decomposition for review before any tool call or code edit occurs, which is exactly where missed goals, duplicated subtasks, and impossible tool assignments are caught. In this scenario, that matters because the refactor spans dozens of files, so the architect needs to approve the structure first rather than discover decomposition errors after execution has begun.
B. Disable tool use and rely on the coordinator's system prompt alone to validate correctness.
A prompt can guide behavior, but it does not expose the actual proposed task breakdown for inspection.
C. Increase the coordinator's temperature so it explores more plan alternatives.
Higher temperature increases variability, not structured validation of completeness, duplication, or feasibility.
D. Run the full workflow end-to-end first and revert with git if the output is wrong.
Post-run rollback mitigates bad outcomes after execution; it does not prevent planning mistakes beforehand.