Cloud Next '26: Google's AI Agent Quality Flywheel Explained

Building agents presents a challenge: ensuring changes truly improve performance. Most teams struggle to connect prompt tweaks with measurable impact in production. The scariest failures are agents that appear functional but miss the user's actual goal. At Cloud Next '26, a three-phase "agent quality flywheel" was introduced: Build & Test, Ship & Monitor, and Learn & Refine. Now, a new developer-facing skill is being added. This skill, installed by a coding agent, drives the quality process on behalf of the developer. It's built on principles Google uses for its own models, with AutoRaters developed with Google DeepMind. The skill focuses on the Build & Test phase, expanding it into five stages. It also runs these stages against production traces. The optimizer and evaluator remain separate; the Gemini Enterprise Agent Platform GenAI evaluation service scores fixes independently. This prevents gaming the metrics. The skill chooses the right metric, runs evaluations, proposes fixes, and compares results. This allows developers to describe a problem in plain language and let the skill handle the evaluation process. This matters because it provides a structured way to improve agent quality and ensures changes are genuinely effective.