Google uses AlphaEvolve and Harbor to optimize agent skills

Skills are rewritten by Gemini, evaluated in real tasks via Harbor, and compared against baseline versions for reliable scoring. After iterations, performance improves by 5%, with the best skill focused on time-out mitigation. Parallel testing reduces noise from model latency. Skills that address specific gaps—like missing business rules—show significant gains, such as +33 points on dev tasks.

Summarised from the original article by an open-source model, then checked automatically: every number and English name must appear in the source. The original article is authoritative.