Upgrading to new AI models presents significant challenges for engineering teams, often involving months of manual testing and evaluation. The rapid pace of AI development, with frequent major model evolutions, exacerbates this issue. Google Cloud's Applied ML team has developed an agentic workflow to drastically reduce model upgrade times from months to hours. This approach is detailed in a blog post, offering three key lessons for accelerating foundation model upgrades.The first lesson emphasizes the importance of hands-on discovery, involving close collaboration with product teams to understand real migration problems and identify prompt optimization needs. The second lesson warns against the rigidity of traditional automation, which struggles with diverse data formats and unique edge cases. The third lesson highlights the success of pivoting to a flexible agent architecture that adapts to specific project requirements.A partner team successfully migrated a video translation and dubbing service to a foundation model using this agentic framework. By providing a dataset and baseline prompt, the system autonomously improved prompt quality, eliminating the need for a custom-tuned model. This process demonstrated the power of prompt engineering over fine-tuning for certain complex tasks.To implement similar migration workflows, teams can leverage Gemini Enterprise Agent Platform and Google Antigravity. Key steps include deploying automated quality evaluators (Autoraters) instead of manual review, building an agentic loop using the Agent Development Kit, and automating orchestration with Antigravity. This shift from manual engineering to intelligent automation reduces tech debt and helps organizations keep pace with AI advancements. This collaborative effort involved numerous contributors across Google.