The Challenge of Scaling AI Agents

As organizations move beyond prototypes, they face significant friction in managing hundreds of disparate agents and tools. Common pain points include poor discoverability of existing assets, redundant tool development, and the difficulty of maintaining a secure, governed posture across legacy and third-party systems. The primary goal for platform teams is to create a "well-lit path" where developers can build, discover, and reuse agents and Model Context Protocol (MCP) servers without creating a fragmented ecosystem.

Core Components for Governed Orchestration

To address these challenges, Google Cloud introduces a framework centered on three pillars:

  • Agent Registry: Acts as a single source of truth for all agents, MCP servers, and models within an organization. It allows developers to search for existing assets and provides code snippets for immediate integration via the ADK.
  • Agent Development Kit (ADK): An open-source framework that simplifies the orchestration of sub-agents and tools. It allows developers to define agents and connect them to registered MCP servers with minimal boilerplate code.
  • Agent Identity & Security: Utilizing SPIFFE-based standards, every agent is assigned a unique identity tied to its lifecycle. This ensures granular permission management, replacing insecure shared service accounts. A secure credentials vault manages API keys and OAuth 2.0 tokens, allowing developers to focus on logic rather than secret management.

Transforming APIs with Spec Boost

Many enterprise APIs lack the rich, accurate specifications required for reliable agent interaction, leading to hallucinations or failures. The Spec Boost feature automates this process:

  1. Generation: It analyzes API logs to generate accurate specifications from scratch.
  2. Optimization: It uses Gemini to refine existing specs, making them more "agent-readable" and human-readable.
  3. Governance: It allows for diffing and merging production traffic against existing documentation, ensuring that specs remain live and accurate without manual overhead.

Once generated, these APIs can be converted into MCP servers with a single click, making them instantly discoverable and usable by other agents in the registry.

Real-World Application: Multi-Agent Orchestration

Klook, a travel experience platform, implemented this architecture to automate supply research. By adopting a multi-layer orchestrator pattern, they replaced manual, multi-day research workflows with an autonomous agentic system. Their approach uses:

  • Orchestrator Sub-agents: Dedicated agents for specific tasks (e.g., social media analysis, internal booking data evaluation) that run in parallel, preventing context anxiety.
  • Reusability: By publishing agents to the registry, Klook ensures that a tool built for one department (e.g., supply research) can be repurposed for another (e.g., voice of customer analysis).
  • Outcome: The system reduced research time from days to hours, allowing business development teams to focus on high-value merchant negotiations rather than data processing.