Building an agent skill optimizer with AlphaEvolve, Harbor and Gemini 4 Argon

Coding agents such as Google Antigravity and OpenCode rely on manually written instructions to perform tasks like building, testing, and deploying code. These skills often lack measurable effectiveness, as demonstrated by tests on Terminal-Bench 2.0 where baseline skills showed no significant improvement over unskilled execution. The challenge lies in identifying which instructions genuinely enhance agent performance without objective assessment.

To address this, Google developed AlphaEvolve, a system that optimizes agent skills by iteratively rewriting instructions and evaluating performance. Harbor serves as the evaluator, running agents on real tasks to score each skill version. The process compares new skills against a baseline, ensuring reliable feedback before refining instructions further, with observed improvements of 5% after a few iterations.

AlphaEvolve treats skills as text files (SKILL.md) and uses Gemini to rewrite them between specific markers, while Harbor executes these skills in isolated containers on Google Kubernetes Engine. Each candidate skill runs alongside the baseline on identical tasks to minimize score variance, with evaluator feedback guiding the next iteration. The system prioritizes addressing the most common failure: timeouts.

Early studies revealed practical challenges, including model latency and log storage constraints during long searches. Skills that added specific guidance—rather than generic advice—yielded measurable gains, such as a 33-point improvement on development tasks. Google is now testing AlphaEvolve on enterprise agent harnesses to refine business-specific rules, with promising initial results in a fictional banking case study.