Google releases Agents CLI eval loop for agent quality hill-climbing

It starts with a small test suite of four scenarios, achieving a mean score of 4.0 with standard deviation 1.41. A prompt error reduces performance to 2.75, with significant degradation in the sf_weather case due to tool invocation failure. eval compare provides per-case differences, exposing behavior shifts undetectable by unit tests and enabling precise regression diagnosis.

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.