Day 2 of Google Cloud Next focused on the developer keynote, highlighting the Gemini Enterprise Agent Platform. This platform, an evolution of Vertex AI, is designed to create autonomous agents that assist users and perform tasks independently.
Gemini Enterprise Agent Platform Overview
Brad Calder, President of GCP and SRE, introduced the platform's capabilities, which include tools for building, scaling, governing, and optimizing agents. The keynote featured a marathon simulation project, illustrating how agents can aid in planning and problem-solving.
Building the Agent
Mofi Rahman demonstrated how the Agent Development Kit (ADK) and Google Cloud's Model Context Protocol (MCP) empower the planner agent with necessary instructions and tools. By the end of the demo, a visually appealing route for the marathon was generated.
Evaluating the Agent
Ivan Nardini and Casey West showcased methods to evaluate the agent's performance and develop a user interface (UI). They introduced the Agent-to-User Interface (A2UI), which streamlines UI creation, and the Agent Registry, which tracks deployed agents.
Enhancing Agent Memory
Calder discussed the importance of creating agents that learn and improve over time. The platform includes features like Agent Platform Sessions and Memory Bank to facilitate knowledge retention and enhance simulation accuracy.
Debugging and Scaling Agents
Megan O’Keefe highlighted the challenges of debugging autonomous agents. She demonstrated how to use Gemini Cloud Assist to identify and resolve issues within the agent system. Bobby Allen then explained how to scale agents using Google Kubernetes Engine (GKE) and migrate to a high-performance file system.
Security and Governance
Emma Twersky emphasized the need for security measures in agent development. Ankur Kotwal presented the Agent Identity and Agent Gateway features, which ensure secure agent actions through IAM policies. Yinon Costica from Wiz introduced tools for scanning agent code and suggesting security remediations.
Conclusion
The keynote concluded with the announcement that all solutions and demos are available as source code on GitHub, providing developers with resources to kickstart their projects.
