Google Cloud shares 7 rules for self-improving AI agent loops, launches agents-cli

Google Cloud has published a set of seven rules for building self-improving agent loops, aimed at AI engineers developing autonomous systems. The guidelines emerged from the team's experience building agents-cli, an open-source command-line tool for creating AI agents on Google Cloud. A central warning in the guidance is that automated loops optimize whatever metric they are given, with no built-in ability to judge whether that metric reflects genuine quality. Google Cloud recommends that engineers define custom, versioned metrics and use a separate grading step to prevent agents from evaluating their own outputs. The release highlights a broader challenge: as agent loops become more automated, the choice of measurement standard matters more than the prompt, code, or agent behavior itself.