Build connected AI: Orchestrate tools and agents with registries and ADK

Key Concepts

• Agent Platform: A Google Cloud ecosystem for building, managing, and governing AI agents at scale.
• Agent Registry: A centralized "single source of truth" for discovering and managing agents, MCP servers, and models.
• MCP (Model Context Protocol): A standard protocol for connecting AI agents to data sources and tools.
• ADK (Agent Development Kit): An open-source framework for building and orchestrating agents.
• Spec Boost: A tool that uses Gemini to generate, optimize, and maintain API specifications from logs.
• Agent Identity: A SPIFFE-based security standard providing unique, lifecycle-bound identities to agents for granular access control.
• Orchestrator Sub-agent Pattern: A design pattern where a "lead" agent delegates tasks to specialized sub-agents to manage context and parallel processing.

---

1. Challenges in Scaling AI Agents

The speakers identify a shift in the industry: moving from building individual agents to managing hundreds of agents within an enterprise. Key challenges include:

• Discoverability: Developers struggle to find existing tools or agents, leading to redundant development (e.g., multiple teams building different servers to query the same BigQuery dataset).
• Governance & Security: Ensuring only authorized agents access specific tools and managing credentials without leaking secrets.
• API Readiness: Many enterprise APIs lack the rich, accurate specifications required for LLMs to function without hallucinating.

2. The Agent Platform Framework

The platform provides a structured approach to these challenges:

• Agent Registry: Acts as a catalog for all organizational assets. It allows developers to search for existing MCP servers and provides code snippets for immediate integration via ADK.
• Agent Identity & Vault: Replaces shared service accounts with unique identities tied to the agent's lifecycle. It includes a Secure Credentials Vault to manage OAuth 2.0 and API keys, allowing developers to focus on logic rather than secret management.
• Spec Boost: Addresses the "Day 2" problem of poor documentation. It analyzes API logs to generate or update OpenAPI specifications, ensuring they are "agent-readable." It includes a human-in-the-loop "diff" process to merge new findings with existing documentation.

3. Real-World Application: Klook’s Supply Research Agent

Min Zhu (Director of Engineering at Klook) presented a case study on automating supply research for their travel platform.

• The Problem: Manual research was slow, failed to scale, and missed "middle-tail" opportunities.
• The Solution: A multi-agent system using an Orchestrator Sub-agent Pattern.
  • Lead Agent: Acts as a project manager, planning and delegating tasks.
  • Specialized Agents: Individual agents analyze specific social platforms or internal booking data in parallel.
  • Outcome: Tasks that previously took days are now completed in hours, allowing the Business Development team to focus on high-value merchant negotiations.

4. Methodologies and Frameworks

• Orchestration: The team advocates for using ADK to wire agents together. For complex workflows, they recommend the "Orchestrator Sub-agent" pattern to prevent "context anxiety" in LLMs.
• Governance Workflow:
  1. API Hub: Centralizes API specs.
  2. Spec Boost: Optimizes specs for LLM consumption.
  3. MCP Conversion: One-click conversion of APIs into MCP servers.
  4. Registry: Publishing the MCP to the Agent Registry for organizational discovery.

5. Notable Quotes

• "The challenge is now how to do it at scale because organizations have hundreds of agents that do things." — Michael, Product Manager.
• "The future of the documentation should be it gets live updated as you publish update your APIs." — Mike, Product Manager.
• "The lead agent plans, delegates, monitors, and adjusts like a project manager." — Min Zhu, Klook.

6. Synthesis and Conclusion

The session emphasizes that the next phase of enterprise AI is governance and connectivity. By moving away from ad-hoc agent development toward a centralized platform (Agent Registry, ADK, and Spec Boost), organizations can eliminate redundant work, improve security through SPIFFE-based identities, and enable complex, multi-agent workflows. The Klook case study serves as a practical validation that this architecture successfully transitions manual, time-intensive business processes into autonomous, scalable AI-driven workflows.