Google Cloud says its Applied ML team has built an agent-based workflow that upgrades AI models in hours rather than months. The system is intended to reduce the manual work involved in moving products to newer model versions.
Engineering teams often face lengthy testing and evaluation when shifting from one foundation model to another or moving to a newer checkpoint within the same model family. That pace has become harder to manage as updates arrive more frequently, with six major model evolutions since 2023 leading to Gemini 3.5.
The workflow is designed to help teams test and refine prompts during a model migration instead of relying on a fixed automation script. Google Cloud says the system emerged after its engineers worked directly with internal product teams on migration problems and found that standard automation struggled with varying data formats and edge cases.
Three lessons
Early work with teams produced prompt-optimisation guidelines that were later turned into a standard automated workflow. That approach delivered some initial gains, but the Applied ML team concluded that it was too rigid for broader use.
It then rebuilt the tool around a flexible agent architecture. In that setup, the agent adapts to the needs of a specific project, analyses data, and tests prompts dynamically during the migration process.
Google Cloud linked the work to two of its products: Gemini Enterprise Agent Platform, which is intended for building and managing agents, and Google Antigravity, which is used for AI coding and orchestration of agent workflows.
Internal use
One internal partner team cited by Google Cloud manages video translation and dubbing services. It needed translated text to be rewritten so the spoken duration matched the pacing of the original video without changing the meaning.
That requirement had historically forced the team to maintain a fine-tuned model. The goal was to move the service to a newer standard foundation model and rely on prompt engineering instead of a custom model stack.
Using the workflow, the team supplied a ground-truth dataset and a baseline prompt. The system then adjusted and improved prompt quality on its own, allowing the service to move away from the custom setup.
Migration process
Other engineering teams could apply the same approach to model upgrades by replacing manual review with model-based automated rating systems, creating an agent loop to test and refine prompts, and using orchestration tools to automate coding and reporting tasks.
That marks a shift in how software teams frame model migration. Rather than treating it as a line-by-line engineering exercise, Google Cloud presents it as a repeatable workflow managed by software agents.
For businesses using large language models, the broader issue is the cost of staying current. New releases can offer better performance or lower operating costs, but each change can trigger quality checks, prompt revisions, and regression testing across products already running in production.
For companies with several AI features, that work can create a growing maintenance burden. A tool that shortens migration cycles could make it easier to adopt newer models more often, especially for teams trying to avoid building and maintaining bespoke models for narrow tasks.
Google Cloud says the project was a joint effort across Google, with contributions from 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, and Zhen Lin.