--- title: "Decoupling AI Agents: Google Introduces the Agent-to-Agent Protocol" type: "News" locale: "zh-HK" url: "https://longbridge.com/zh-HK/news/271835799.md" description: "Google has introduced the Agent-to-Agent (A2A) protocol and the Agent Development Kit (ADK) to address the complexities of integrating disparate AI systems. The A2A protocol acts as a universal interface, allowing agents built in different languages to communicate seamlessly. The AgentCard, a machine-readable JSON file, specifies each agent's capabilities, enabling effective communication without internal code dependencies. The ADK introduces an Orchestrator agent to manage state and execution control, ensuring lightweight, scalable worker agents. This architecture promotes fault tolerance and simplifies workflow management, enhancing the development of robust AI systems." datetime: "2026-01-07T20:50:50.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/271835799.md) - [en](https://longbridge.com/en/news/271835799.md) - [zh-HK](https://longbridge.com/zh-HK/news/271835799.md) --- > 支持的語言: [简体中文](https://longbridge.com/zh-CN/news/271835799.md) | [English](https://longbridge.com/en/news/271835799.md) # Decoupling AI Agents: Google Introduces the Agent-to-Agent Protocol The primary friction point in scaling modern AI systems is not computational power or model quality, but the sheer complexity of integration. When disparate AI services—built across different frameworks, languages, and API schemas—must communicate, developers often spend more time writing “glue code” than actual AI logic, resulting in brittle, monolithic architectures. Google Developer Advocate Amit Maraj, in a recent presentation on building multi-agent systems, introduced a foundational solution to this problem: the Agent-to-Agent (A2A) protocol, paired with the Agent Development Kit (ADK), providing a standardized, microservice-like framework for agent interoperability. Maraj spoke about the necessity of a universal interface, likening the current state of AI integration to wrestling with incompatible travel adapters. The A2A protocol aims to be that universal adapter, ensuring that agents, regardless of their underlying stack—Python, Node, or Go—can communicate seamlessly via a standard HTTP interface. This standardization hinges on a core concept: the AgentCard, a machine-readable JSON file hosted at \`/.well-known/agent.json\`. This AgentCard serves as the agent’s public contract. It specifies the agent’s name, description, version, and, critically, its capabilities, including input and output schemas. Maraj notes that the AgentCard is “kind of like our agents’ dating profile,” providing all necessary information for another agent or orchestrator to interact with it intelligently. By defining these schemas explicitly, agents can communicate effectively without needing to import or understand the internal code dependencies of their peers. This is analogous to how modern microservices use OpenAPI specifications (Swagger) to define their interfaces, but tailored specifically for the dynamic, prompt-driven nature of large language models (LLMs). The immediate benefit of this approach is architectural decoupling. If a Researcher agent is built using a super-fast, cheap model in Python, and a subsequent Content Writer agent uses a massive, expensive reasoning model in Go, they can exchange data without caring about each other’s implementation details. They simply speak A2A. Maraj emphasized that this enables true polyglot persistence and architectural flexibility: “You could even rewrite the Researcher in Go tomorrow, keep the same A2A interface, and the rest of your system wouldn’t even blink. That’s true microservice architecture, but for AI.” This separation of concerns is paramount for enterprises looking to iterate quickly, swap models based on cost or performance, and maintain resilience against technological churn. Interoperability through A2A is only half the battle; complex AI workflows require robust state management and execution control. This is where the Agent Development Kit (ADK) introduces a crucial architectural pattern centered on the Orchestrator agent. Instead of having every worker agent maintain its own memory of past steps and context—a method that is inherently fragile and resource-intensive—the ADK enforces a pattern where a single Orchestrator manages the overall state. The Orchestrator functions as a “Project Manager” maintaining the “Master Clipboard” of state. This design ensures that specialized worker agents, such as the Researcher or the Judge, remain lightweight, scalable, and crash-resistant. Worker agents, being stateless, receive instructions and context from the Orchestrator, execute a specific task, and return the output to the master clipboard before terminating or awaiting the next instruction. This architecture is inherently fault-tolerant. As Maraj points out, “If a worker bot crashes and restarts, it’s fine. It didn’t lose anything important because it wasn’t holding anything important.” The Orchestrator retains the complete history and context, allowing failed steps to be easily retried or re-assigned without losing critical progress. This is a fundamental divergence from earlier agent frameworks that often required persistent, stateful worker processes, making deployment and scaling difficult. The ADK further simplifies the orchestration layer by providing standardized patterns for managing workflow, such as the Sequential Agent (do step A, then B) and the Loop Agent (iterate step A until a condition is met). These patterns abstract away the complex logic of task routing and state persistence, allowing developers to focus purely on defining the workflow and the data structure of the state—the “clipboard”—that flows between agents. The emphasis, therefore, shifts from writing complex integration code to defining the precise data contracts (via A2A) and the high-level execution flow (via ADK patterns). For sophisticated AI ecosystems, defining what data needs to be saved—such as research findings and judge feedback—becomes the foundational step before any code is written. This disciplined approach promises to elevate AI agent construction from experimental scripting to industrial-grade, distributed systems engineering. ### 相關股票 - [Alphabet - C (GOOG.US)](https://longbridge.com/zh-HK/quote/GOOG.US.md) - [Alphabet (GOOGL.US)](https://longbridge.com/zh-HK/quote/GOOGL.US.md) - [Direxion Daily Googl Bear 1x Shares (GGLS.US)](https://longbridge.com/zh-HK/quote/GGLS.US.md) - [Direxion Daily GOOGL Bull 2X Shares (GGLL.US)](https://longbridge.com/zh-HK/quote/GGLL.US.md) ## 相關資訊與研究 - [Google in Talks With Poolside to Revive Data Center Project](https://longbridge.com/zh-HK/news/281507610.md) - [Google's Nobel Prize-Winning AI Chief: 'We Don't Feel Any Pressure' While OpenAI Burns $14 Billion](https://longbridge.com/zh-HK/news/281220147.md) - [Experts urged Google to prohibit AI Videos targeting young YouTube users](https://longbridge.com/zh-HK/news/281386661.md) - [Google Lets Gmail Users Change Old Email Addresses Without Losing Accounts](https://longbridge.com/zh-HK/news/281579693.md) - [Google has been developing screenless Fitbit band and plans to release it later this year - Bloomberg News](https://longbridge.com/zh-HK/news/281269407.md)