Google demos Agents CLI eval workflow for agent quality

On November 2, 2026, Google Developers Blog published an overview of the Agents CLI evaluation loop, with the goal of systematically improving agent performance through continuous testing and grading. The workflow has four parts: eval generate runs the agent on a dataset, eval grade scores the resulting traces, eval compare places two runs side by side, and eval optimize performs hill-climbing updates on prompts. The post also explains how to add nightly scheduled runs together with threshold-based alerting to track quality over time. The walkthrough covers a first run through to regression tracking, model-backend experiments, and automated alerting, including catching prompt regressions via side-by-side trace comparisons, evaluating model backend swaps, and automating evaluation passes. The material available consists of this single post; there are no other sources, follow-up developments, or divergent accounts to report.