Google Cloud、基盤モデルのアップグレード加速に関する教訓を共有

Have you run into problems migrating your products from one model to the next?

Upgrading to the latest AI models is rarely simple. For engineering teams, model updates whether migrating to an entirely new model or updating to a newer checkpoint within the same model family, like moving from an earlier Gemini version to Gemini 3.5 — often require a slow and costly process of testing, proving quality, and manually evaluating new responses. For most engineering teams, upgrading to a new model checkpoint means months of manual toil to verify performance. And the industry is moving at breakneck pace – since 2023, we’ve announced six major model evolutions, bringing us to Gemini 3.5 today.

Our team at Google Cloud, Applied ML, has a goal to deliver transformative infrastructure and services that benefit both Google and our customers globally. As part of that, our team built an agentic workflow that completes model upgrades in hours instead of months.

In this blog, we’ll show you our approach and three lessons you can apply to accelerate your own foundation model upgrades using Gemini Enterprise Agent Platform — our new, comprehensive platform to build, scale, govern, and optimize agents – and Google Antigravity, our primary solution for developers using AI for coding and agent orchestration.

Three lessons in building a flexible agent system

To support different team needs, we had to rethink traditional automation and learned three key lessons along the way:

• Lesson 1: Start with hands-on discovery. First, our engineers worked closely with product teams on real migration problems. This hands-on work helped us identify complex requirements and build our first guidelines for prompt optimization.

• Lesson 2: Beware the rigidity of traditional automation. We turned these guidelines into a standard, automated workflow. While this version gave us some quick wins, we soon found that traditional automation was too rigid to handle different data formats and unique edge cases.

• Lesson 3: Pivot to a flexible agent architecture. The real progress came when we rebuilt the tool using a flexible agent. Instead of forcing teams into a rigid process, the agent adapted to specific project needs, helping analyze data and test prompts dynamically with a high degree of adaptability.

How our partner teams cut migration time while boosting quality

Our partner team, which manages video translation and dubbing services, had an interesting challenge: their workflow required rewriting translated text so that the spoken duration matched the original video's pacing exactly, without altering the meaning. Historically, this strict constraint required maintaining a fine-tuned model. Their goal was to migrate to the latest out-of-the-box foundation model, guided purely by prompt engineering.

Using this agentic framework, the team provided their ground-truth dataset and baseline prompt. The system autonomously hill-climbed the prompt quality, migrating the service away from the custom stack

Make your own migration workflow with Agent Platform and Google Antigravity

These learnings can be applied by any engineering team looking to accelerate their own model upgrades. If your organization is struggling to keep pace with new foundational models, replacing manual toil with intelligent automation requires treating migration as an agentic workflow.

To build your own automated migration pipeline, follow these steps:

• Deploy Autoraters: Pivot from manual human review to model-based Autoraters to evaluate the quality of a new checkpoint at scale and in a fraction of the time.

• Build an agentic loop: You can use the Agent Development Kit within Gemini Enterprise Agent Platform to create your agent.

• Automate the orchestration: To make the process even easier, leverage Antigravity to automate the underlying coding and agent orchestration and add in features such as loss reporting or headroom reports.

By shifting away from a manual, line-by-line engineering task, organizations can reduce infrastructural tech debt and confidently keep pace with the frontier of AI.

This work is the result of collaboration across Google. We thank key contributors: Anthony Green, Chris Lamb, Chungyen Li, Connie Huang, Elaine Han, Elena Erbiceanu Tener, Eugene Ie, Francesca Ciacchella, Igor Karpov, Jeanie Jung, Jose Menendez, Kiam Choo, Lina Sanders-Self, Longfei Shen, Martin Nikoltchev, Mason Ng, Matt Mancini, Paul Zhou, Pedram Oskouie, Samuel Smith, Tom Lawrie, Ye Tian, Zhen Lin