Google Agent Quality Flywheel [Coding Skill]
Bottom Line
Shipping agents without a closed loop is vibe-checking. Google’s flywheel skill forces Build & Test → Ship & Monitor → Learn & Refine with AutoRaters so prompt tweaks cannot silently regress ten other cases.
Key Takeaways
- ›The scary failures are quiet: confident plans that miss the user’s real goal.
- ›Most teams have evals and tweak prompts; few connect the two with enough discipline.
- ›Cloud Next ’26 framed agent quality as a three-phase flywheel; this post adds a coding-agent skill path.
- ›AutoRaters are aligned with principles used on first-party Google/DeepMind agents.
- ›Install the skill so your coding agent drives eval loops on your behalf.
Google’s “Driving the Agent Quality Flywheel from Your Coding Agent” (June 30, 2026) is a practical ops manual for agent quality.
The gap that keeps PMs awake
You ship an agent. It works. You tweak a prompt for one complaint. It looks better on three examples. Did you break ten others? That gap — between “looks better on a few examples” and “actually better in production” — is the daily reality of agent teams.
The flywheel
| Phase | Job |
|---|---|
| Build & Test | Fast iteration with evals and AutoRaters |
| Ship & Monitor | Detect silent goal failures in production |
| Learn & Refine | Feed failures back into the eval set |
The coding-agent skill centers on Build & Test and expands it into five concrete stages so a coding agent can drive the loop instead of humans hand-copying results into spreadsheets.
Engineering adoption checklist
- Codify 20–50 golden multi-turn tasks from real tickets
- Run AutoRater-style graders on state, not free-text vibes
- Block prompt merges without eval delta reports
- Promote silent failures (wrong goal, confident tone) into the suite