What's new in Google Cloud's agent platform

Key Concepts

• Agent Platform: A vertically integrated Google Cloud stack for building, scaling, governing, and optimizing AI agents.
• ADK (Agent Development Kit): The foundational framework for building agents, now featuring graph-based orchestration and multi-agent collaboration.
• Agent Engine: The runtime environment for deploying agents, supporting long-duration tasks (up to 7 days) and bidirectional streaming.
• Agent Registry & Gateway: Centralized control planes for managing agent identity, permissions, and traffic security.
• Model Armor: A security layer designed to block prompt injection and prevent data leaks.
• MCP (Model Context Protocol): A standard protocol used for interoperability between agents, tools, and external data sources.
• Auto-Raters: Built-in scoring mechanisms for automated offline and online evaluation of agent performance.

---

1. The Evolution of the Agent Platform

The presentation highlights a shift from simple RAG (Retrieval-Augmented Generation) chatbots to sophisticated, autonomous agents that act as the "engine of business." The speakers compare traditional software to a reliable "Toyota Camry," while characterizing agents as "Formula One race cars"—highly powerful but requiring robust guardrails to operate safely.

The platform is designed to eliminate the "handoff tax" between developers and end-users by providing a unified experience that connects Gemini Enterprise (the user-facing interface) with the Agent Platform (the developer backend).

2. Core Pillars of the Platform

The platform is structured around four primary pillars:

• Building (ADK):
  • Graph-based Orchestration: Allows modeling of complex business processes with branches, loops, and conditional paths.
  • Agent Collaboration: Enables the decomposition of complex tasks across specialized agents rather than relying on a single monolithic agent.
  • Native Skill Definitions: Skills can be defined once and reused across different agents.
• Scaling (Agent Engine):
  • Runtime Improvements: Features sub-second cold starts and "bring your own container" support.
  • Memory & Sessions: Agents now possess persistent memory, allowing them to retain context, user preferences, and learning over time.
  • Sandbox Environments: Includes code execution sandboxes and GUI interaction capabilities, enabling agents to perform real-world tasks.
• Governance (End-to-End Security):
  • Agent Identity: Granular, auditable permissions for every agent.
  • Agent Registry: A centralized catalog to track all agents, their tools, and their access levels.
  • Proactive Threat Management: Integrated monitoring to detect and mitigate risks like prompt injection.
• Optimization:
  • Auto-Raters: Leverages DeepMind research to provide out-of-the-box evaluation metrics.
  • Prompt Optimization: A single pane of glass for continuous feedback loops, allowing developers to refine agents based on performance data.

3. Real-World Application: L’Oréal Case Study

L’Oréal shared their journey of scaling AI from a simple internal chatbot ("L’Oréal GPT") to a comprehensive Beauty Tech Platform.

• The Shift: They moved from a "make" strategy (building custom services) to a "buy" strategy (leveraging Google Cloud’s Agent Platform) to focus on business value rather than infrastructure maintenance.
• Scale: They currently support 41,000 weekly users and 30,000 "zero-code" agents developed by employees.
• Human Stack: L’Oréal emphasizes that technology alone is insufficient. They implemented a "human stack" consisting of:
  • Center of Excellence: Experts providing starter packs and support.
  • The Lineup: A community of "AI champions" across the organization who bridge the gap between IT and business units.
  • Governance Task Force: Ensuring all agents comply with cybersecurity, legal, and data privacy standards.

4. Notable Quotes

• "With agents, we're building race cars... but you can't just run a race car in YOLO mode." — Iman Khan, on the necessity of guardrails.
• "The key shift is that we went from building services for everyone to empowering everyone to build." — Gauthier, L’Oréal, on scaling AI adoption.

5. Synthesis and Conclusion

The Google Agent Platform represents a transition toward enterprise-grade, autonomous AI. By providing a unified stack that handles the "plumbing"—from identity and security to evaluation and memory—Google enables organizations to move from prototyping to production-ready workflows. The success of L’Oréal demonstrates that the most effective way to scale AI is to combine a robust, standardized technical platform with a human-centric community that empowers non-technical employees to build and deploy their own solutions.