Coding agents are increasingly used for software engineering tasks but require effective instructions. Manually written agent skills are difficult to refine as their effectiveness is hard to assess. Initial tests with OpenCode and Gemini 3.8 Flash on Terminal-Bench 2.0 showed inconclusive results with a baseline skill. To improve performance, AlphaEvolve was employed to optimize skills by iteratively proposing new versions, scoring them with an evaluator, and retaining the best. Harbor was used as the evaluator, running agents on real tasks in isolated containers. This process treats agent skills like programs that AlphaEvolve can rewrite. The evaluator in Harbor compares candidate skills against a baseline skill simultaneously on the same tasks to ensure trustworthy scoring. A portion of tasks is reserved to test the final optimized skill on unseen data. AlphaEvolve focused on addressing the most common failure: running out of time. The system achieved a 5% improvement in a few iterations and a significant 33-point increase on specific development tasks. Challenges encountered included planning for slow model responses and log storage during long searches. Future work involves testing AlphaEvolve on enterprise agent harnesses to incorporate business rules.