Upgrading to new foundation models is rarely simple. Engineering teams often face months of manual testing, prompt optimization, and quality verification to move from one checkpoint to the next — even within the same model family. Google Cloud‘s Applied ML team notes that the industry has accelerated dramatically, with six major model evolutions announced since 2023, yet most teams still rely on slow, costly processes to validate each upgrade. The core tension is clear: the frontier moves fast, but the operational burden of keeping up remains high.

To solve this, the team built an agentic workflow using the Gemini Enterprise Agent Platform and Google Antigravity. They started with hands-on discovery alongside product teams, then attempted a rigid automated pipeline — but found it too brittle for varied data formats and edge cases. The breakthrough came when they pivoted to a flexible agent architecture that adapts to each project’s needs, dynamically analyzing data and hill-climbing prompt quality. A partner team managing video translation and dubbing used this approach to migrate from a fine-tuned model to a pure prompt-engineering stack, cutting migration time from months to hours while maintaining strict timing constraints.

The key takeaway for builders is to treat model migration as an agentic workflow rather than a manual line-by-line engineering task. Deploy Autoraters to replace human review with scalable, model-based evaluation. Build an agentic loop using the Agent Development Kit, and automate orchestration with Antigravity to add features like loss reporting and headroom reports. This pattern reduces infrastructural tech debt and lets teams confidently keep pace with the rapid release of new foundation models.