Accelerate Foundation Model Upgrades with Agentic Workflows

Google Cloud's Applied ML team has built an agentic workflow designed to complete foundation model upgrades in hours, a significant improvement over the traditional multi-month manual process. This new approach aims to provide transformative infrastructure and services for both Google and its customers.

Engineering teams can accelerate their own model upgrades by adopting an agentic workflow. This involves deploying model-based 'Autoraters' for quality evaluation, building an agentic loop using the Agent Development Kit within Gemini Enterprise Agent Platform, and automating orchestration with Antigravity for features like loss and headroom reporting.

The team learned three key lessons during development: 1) Start with hands-on discovery by working closely with product teams to identify complex requirements and establish prompt optimization guidelines. 2) Be aware of the rigidity of traditional automation, which may struggle with diverse data formats and edge cases. 3) Pivot to a flexible agent architecture that can adapt to specific project needs for dynamic data analysis and prompt testing.

A partner team managing video translation and dubbing services successfully migrated their workflow to a latest out-of-the-box foundation model using the agentic framework. By autonomously optimizing prompts based on ground-truth data and a baseline prompt, they moved away from a custom fine-tuned model stack.

Upgrading to newer AI model checkpoints, such as moving from Gemini to Gemini 3.5, traditionally involves lengthy and costly manual testing, quality assurance, and response evaluation, often taking months. The rapid pace of AI development, with multiple major model evolutions announced since 2023, exacerbates this challenge for engineering teams.