Google Cloud sets out five rules for evaluating agent skills

[Google Cloud] Google Cloud said its team has spent recent weeks measuring and optimizing its Google Cloud Developer Plugin, running dozens of scenarios with Gemini, Antigravity and other agents, using Claude Sonnet 5 on the Gemini platform inside isolated Harbor pods on GKE Autopilot, and contributing Kubernetes and Gemini platform capabilities upstream to NVIDIA's open-source evaluator. The team outlined five engineering rules for verifying whether a skill is worth its tokens. It said trigger collisions and security risks should be gated before containers are spun up, because coding agents do not read full SKILL.md bodies at turn zero and the router only inspects the short name and description in the YAML frontmatter. It also warned against trusting evaluations run on a developer laptop, where cached credentials and exported project variables contaminate the baseline, and said teams should measure paired WITH versus W/OUT ablation at Pass@3 rather than rely on impressions. According to the team, more than a third of benchmark tasks in its retrieval study produced different scores across consecutive runs of the exact same code and prompt.