Google Cloud builds agent skill optimizer with AlphaEvolve, Harbor and Gemini

[Google Cloud] Google Cloud said its team built an agent skill optimizer using AlphaEvolve, Harbor and Gemini, treating the text of SKILL.md skills as an evolvable program in which Gemini proposes new versions, an evaluator scores them and the search keeps the higher-scoring version. According to Google Cloud, on the public Terminal-Bench 2.0 benchmark with OpenCode running Gemini 3.8 Flash, the agent passed about two-thirds of its attempts with the baseline skill, but results were about the same without the skill; after a few AlphaEvolve iterations, the team saw improvements of 5%. Google Cloud said Harbor runs real coding agents on Google Kubernetes Engine with one pod per trial, and a single search ran thousands of agent trials, with the evaluator returning pass and failure scores plus short notes on failed or timed-out tasks that Gemini uses for the next version. Google Cloud also said that in an early run on the same benchmark its model's latency explained much of the variance in candidate scores, so each candidate runs next to the baseline skill on the same tasks at the same time and scores are the difference; about a third of the search's trials hit their time limit, and the best skill is built around that limit.