# How A2A and MCP work together: five integration patterns for building multi-agent systems
**作者**: Google Cloud Tech
**日期**: 2026-04-24T06:45:57.000Z
**来源**: [https://x.com/GoogleCloudTech/status/2047567704807346675](https://x.com/GoogleCloudTech/status/2047567704807346675)
---
No organization will build every agent it needs from scratch. The real value comes from discovering agents built by different teams, in different languages, across different organizations. That's what A2A and MCP enable: A2A for agent-to-agent communication. MCP for agent-to-tool communication.
By @addyosmani and @Saboo_Shubham_
At Cloud Next 26, we shipped the infrastructure that makes this integration practical at enterprise scale. Here are five integration patterns for building multi-agent systems with both protocols.
## Pattern 1: Agent Card Discovery
Before your agent can delegate to another agent, it needs to know that agent exists, what it can do, and how to communicate with it.
A2A solves this with Agent Cards. Every A2A-compatible agent publishes a JSON document at a well-known URL that describes its capabilities, authentication requirements, and rate limits. Think of it as an OpenAPI spec, but designed for agent-to-agent interaction rather than client-server communication.
In Agent Development Kit (ADK), exposing your agent as an A2A server is a single configuration. It automatically generates the Agent Card from your agent's definition. Consuming a remote agent is equally straightforward through the RemoteA2aAgent component, which handles authentication, data serialization, error handling, and result streaming. From the coordinator's perspective, a remote agent built by another team in another language feels like a local one.
Agent Registry amplifies this. When you register agents in the central registry, other agents across your organization can discover them without knowing specific URLs. The registry becomes the service mesh for your agent ecosystem: a single place to find what capabilities are available, who maintains them, and how to access them.
## Pattern 2: Delegated Specialization
The most common integration pattern is delegation: your agent encounters a task outside its expertise and hands it to a specialist.
This pattern builds on the Coordinator-Dispatcher architecture, but extends it across team and framework boundaries. The specialist doesn't need to be built with ADK. It doesn't need to be written in the same language. It doesn't need to be maintained by the same team. It just needs to speak A2A.
Consider a customer onboarding workflow that spans five teams and four languages. The coordinator (your team, Python) delegates identity verification to the security team's Go agent, credit assessment to the risk team's Java agent, account provisioning to the platform team's Go agent, compliance documentation to the legal team's Python agent, and welcome communication to the marketing team's TypeScript agent.
Each specialist is owned by a different team, maintained independently, and deployed on its own schedule. The coordinator doesn't know or care about internal implementations. It only knows their Agent Card capabilities and the A2A protocol.
When the security team updates their verification agent to support new document types, no changes are needed in the coordinator. When the risk team improves their credit models, the coordinator automatically gets better assessments. Each team iterates independently, and the overall system improves. This is the same principle that made microservices successful: independent deployment, independent scaling, independent evolution. A2A provides the service contract that makes it work for agents.
## Pattern 3: Tool Bridge (Model Context Protocol)
While A2A connects agents to agents, MCP connects agents to tools and data. The Tool Bridge pattern uses MCP to give your agents secure access to data sources, APIs, and enterprise systems without building custom connectors for each one.
Without MCP, every data connection requires its own custom integration code. Your agent needs a Python connector for one REST API, a different connector for another database, yet another for a legacy system. Three data sources means three connectors to build, test, and maintain. MCP eliminates this by providing a single protocol that any tool can implement.
The ADK integrations ecosystem provides more than 60 ready-to-use tools and integrations, such as GitHub, Notion, Hugging Face, AgentOps, and Stripe. These allow your agent to connect instantly through simple configuration, bypassing the need for lengthy custom development.
For database connectivity specifically, MCP Toolbox for Databases connects over 30 different data sources to your agents through a single MCP interface. And Apigee API Hub takes this further: if you have existing REST APIs documented in Apigee, you can turn them into agent-accessible tools automatically. Your agents access data through the same governance layer that manages your existing API traffic. Rate limiting, authentication, logging, and access control are handled by infrastructure you already manage.
ADK supports MCP natively. When you add an MCP tool to your agent, the agent discovers the tool's capabilities, invokes it with structured parameters, and receives structured responses. From the agent's perspective, an MCP tool through Stripe looks the same as an MCP tool through BigQuery. The protocol is the interface. The backend is interchangeable.
## Pattern 4: Cross-Organization Federation
The most ambitious pattern involves agents from different organizations collaborating on shared tasks. This is where A2A's open protocol design becomes essential, because you can't assume the other organization uses the same framework, language, or cloud provider.
Agent Gallery in Gemini Enterprise makes this practical. Over 100 validated partner agents from Adobe, ServiceNow, Workday, Salesforce, and others are available directly within the platform. Every partner agent has been validated by Google Cloud for security and interoperability.
The critical property is that each organization maintains its own governance. Your Agent Gateway policies control what data your agents can share with the partner agent, what actions they can take based on the partner's responses, and what information they can request. The partner agent operates under its own security model, validated by Google Cloud's integration requirements.
Consider a multi-agent payment workflow where your agent needs to coordinate with external merchant agents and financial service agents. Each party maintains its own governance while collaborating on shared transactions through A2A. Your agent doesn't need to understand how the merchant's system works internally. It communicates through the A2A protocol, and each side enforces its own security boundaries independently.
The same principle applies to any cross-organization integration. Your agent doesn't need to understand Salesforce's data model or ServiceNow's internal architecture. It delegates to the partner agent in Agent Gallery, which is maintained by the partner's team, validated by Google Cloud, and governed by your Agent Gateway policies. When the partner updates their platform, their agent is updated. Your system benefits without any changes on your end.
## Pattern 5: Ambient Event Mesh
The final pattern combines A2A with event-driven architecture to create a mesh of ambient agents that react to events continuously in the background, without waiting for user requests.
Batch and Event-Driven Agents in Gemini Enterprise Agent Platform connect directly to BigQuery tables and Pub/Sub streams. When you combine this with A2A, you create a network of agents that listen for events, process them, and delegate to specialist agents as needed - all running continuously in the background.
Each agent in the mesh runs independently. When an event arrives, the receiving agent processes it and decides whether to handle it locally, delegate to a specialist via A2A, or escalate to a human via Mission Control. The mesh is self-organizing - adding a new specialist agent (say, a fraud detection specialist) requires only registering it in Agent Registry and updating the routing logic in the relevant ambient agents.
This pattern is particularly powerful for organizations with high event volumes. A content platform processing millions of uploads per day, a financial institution processing millions of transactions, an e-commerce platform processing millions of customer interactions - all can use ambient agent meshes to handle the long tail of events that require intelligent processing.
The governance story is critical here. Every agent in the mesh has its own identity via Agent Identity. Every tool access is governed by Agent Gateway. Every interaction is traced through Agent Observability. The mesh is not a black box - it is a fully observable, governed network of specialized agents.
## Choosing the Right Pattern
When building multi-agent systems, choose your integration pattern based on the relationship between agents:
Get started today
The interoperability stack is available now:
- A2A protocol: Supported in ADK across Python, TypeScript, Go, and Java
- MCP: Native support in ADK, managed support for GCP databases
- Agent Gallery: 100+ validated partner agents in Gemini Enterprise
- Native Ecosystem Integrations: Plug-and-play connectors for enterprise systems
The A2A + MCP codelab walks through building a multi-agent system with cross-language collaboration: codelabs.developers.google.com/instavibe-adk-multi-agents
ADK samples including multi-agent patterns: github.com/google/adk-samples
The companies building isolated agents will be refactoring in twelve months. The companies building interoperable agent ecosystems will be compounding their advantage every day.
Start building: adk.dev | https://cloud.google.com/products/gemini-enterprise-agent-platform
Build Interoperable Agents: Google for Startups AI Agents Challenge
Isolated agents are a dead end. Join our 6-week global challenge for startups to build, optimize or refactor AI Agents on the Gemini Enterprise Agent Platform. You'll get $500 in cloud credits, full platform access and a shot at the $90,000 prize pool.
Sign up today to start building!
## 相关链接
- [Google Cloud Tech](https://x.com/GoogleCloudTech)
- [@GoogleCloudTech](https://x.com/GoogleCloudTech)
- [6.3K](https://x.com/GoogleCloudTech/status/2047567704807346675/analytics)
- [@addyosmani](https://x.com/@addyosmani)
- [@Saboo_Shubham_](https://x.com/@Saboo_Shubham_)
- [Coordinator-Dispatcher architecture](https://developers.googleblog.com/developers-guide-to-multi-agent-patterns-in-adk/)
- [ADK integrations ecosystem](https://developers.googleblog.com/supercharge-your-ai-agents-adk-integrations-ecosystem/)
- [MCP Toolbox for Databases](https://adk.dev/integrations/mcp-toolbox-for-databases/)
- [codelabs.developers.google.com/instavibe-adk-multi-agents](https://codelabs.developers.google.com/instavibe-adk-multi-agents)
- [github.com/google/adk-samples](http://github.com/google/adk-samples)
- [adk.dev](https://adk.dev/)
- [https://cloud.google.com/products/gemini-enterprise-agent-platform](https://cloud.google.com/products/gemini-enterprise-agent-platform)
- [Sign up today to start building](https://devpost.team/hackathon_guest_invites/4fb181b4-2722-415d-a442-285a57dcaba5?utm_source=twitter&utm_medium=social&utm_campaign=google-for-startups-ai-agents-challenge&utm_content=twitter-post)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [2:45 PM · Apr 24, 2026](https://x.com/GoogleCloudTech/status/2047567704807346675)
- [6,339 Views](https://x.com/GoogleCloudTech/status/2047567704807346675/analytics)
- [View quotes](https://x.com/GoogleCloudTech/status/2047567704807346675/quotes)
---
*导出时间: 2026/4/24 17:32:34*
---
## 中文翻译
# A2A 与 MCP 如何协同工作：构建多智能体系统的五种集成模式
**作者**： Google Cloud Tech
**日期**： 2026-04-24T06:45:57.000Z
**来源**： [https://x.com/GoogleCloudTech/status/2047567704807346675](https://x.com/GoogleCloudTech/status/2047567704807346675)
---
没有任何组织会从零开始构建它所需的所有智能体。真正的价值在于发现由不同团队、使用不同语言、跨越不同组织构建的智能体。这就是 A2A 和 MCP 所实现的：A2A 用于智能体与智能体之间的通信，MCP 用于智能体与工具之间的通信。
作者：@addyosmani 和 @Saboo_Shubham_
在 Cloud Next 26 大会上，我们发布了让这种集成在企业级规模上切实可行的基础设施。以下是使用这两种协议构建多智能体系统的五种集成模式。
## 模式 1：智能体卡片发现
在你的智能体可以将任务委托给另一个智能体之前，它需要知道该智能体的存在、它能做什么，以及如何与它通信。
A2A 通过智能体卡片（Agent Cards）解决这个问题。每个兼容 A2A 的智能体都会在一个众所周知的 URL 上发布一份 JSON 文档，描述其能力、身份验证要求和速率限制。可以把它想象成 OpenAPI 规范，但它是为智能体间交互而设计的，而非客户端-服务器通信。
在智能体开发套件（ADK）中，将你的智能体公开为 A2A 服务器只需一项配置。它会根据智能体的定义自动生成智能体卡片。通过 RemoteA2aAgent 组件调用远程智能体同样简单，它会处理身份验证、数据序列化、错误处理和结果流式传输。从协调者的角度来看，由另一个团队用另一种语言构建的远程智能体，感觉就像本地智能体一样。
智能体注册表进一步放大了这一优势。当你在中央注册表中注册智能体后，整个组织中的其他智能体无需知道具体的 URL 就能发现它们。注册表成为你的智能体生态系统的服务网格：在这里可以一站式了解有哪些能力可用、由谁维护，以及如何访问。
## 模式 2：委托专业化
最常见的集成模式是委托：你的智能体遇到超出其专业范围的任务时，会将其交给专家处理。
这种模式建立在协调者-分发者架构的基础上，但将其扩展到了团队和框架边界之外。专家智能体不需要用 ADK 构建，不需要用相同的语言编写，也不需要由同一个团队维护。它只需要支持 A2A 协议即可。
设想一个横跨五个团队、四种语言的客户入驻工作流。协调者（你的团队，Python）将身份验证委托给安全团队的 Go 智能体，将信用评估委托给风险团队的 Java 智能体，将账户配置委托给平台团队的 Go 智能体，将合规文档委托给法务团队的 Python 智能体，将欢迎沟通委托给市场团队的 TypeScript 智能体。
每个专家智能体由不同的团队负责、独立维护，并按自己的节奏部署。协调者不了解也不关心内部实现。它只了解它们的智能体卡片能力和 A2A 协议。
当安全团队更新其身份验证智能体以支持新的文档类型时，协调者无需做任何更改。当风险团队改进其信用模型时，协调者会自动获得更好的评估结果。每个团队独立迭代，整个系统随之改进。这与微服务成功的原理相同：独立部署、独立扩展、独立演进。A2A 提供了让这一原理在智能体领域生效的服务契约。
## 模式 3：工具桥接（模型上下文协议）
A2A 连接的是智能体与智能体，而 MCP 连接的是智能体与工具和数据。工具桥接模式使用 MCP 让你的智能体安全地访问数据源、API 和企业系统，而无需为每一个系统构建自定义连接器。
没有 MCP，每个数据连接都需要自己的自定义集成代码。你的智能体需要一个用于某个 REST API 的 Python 连接器，一个用于另一个数据库的连接器，还需要一个用于遗留系统的连接器。三个数据源意味着要构建、测试和维护三个连接器。MCP 通过提供任何工具都可以实现的单一协议，消除了这种负担。
ADK 集成生态系统提供 60 多个开箱即用的工具和集成，例如 GitHub、Notion、Hugging Face、AgentOps 和 Stripe。通过简单的配置，你的智能体即可即时连接，省去了冗长的自定义开发。
具体到数据库连接方面，MCP Toolbox for Databases 通过单一 MCP 接口将 30 多种不同的数据源连接到你的智能体。而 Apigee API Hub 则更进一步：如果你在 Apigee 中已有文档化的 REST API，可以自动将它们转换为智能体可访问的工具。你的智能体通过管理现有 API 流量的同一治理层来访问数据。速率限制、身份验证、日志记录和访问控制都由你已经在管理的基础设施处理。
ADK 原生支持 MCP。当你向智能体添加 MCP 工具时，智能体会发现该工具的能力、使用结构化参数调用它，并接收结构化响应。从智能体的角度来看，通过 Stripe 的 MCP 工具与通过 BigQuery 的 MCP 工具看起来是一样的。协议即接口。后端可以互换。
## 模式 4：跨组织联邦
最具雄心的模式涉及来自不同组织的智能体协作完成共享任务。这正是 A2A 开放协议设计变得至关重要的地方，因为你不能假设对方组织使用相同的框架、语言或云服务提供商。
Gemini Enterprise 中的智能体画廊（Agent Gallery）让这一切变得切实可行。来自 Adobe、ServiceNow、Workday、Salesforce 等公司的 100 多个经过验证的合作伙伴智能体可直接在平台内使用。每个合作伙伴智能体都经过了 Google Cloud 的安全性和互操作性验证。
关键特性在于每个组织都保持自己的治理体系。你的智能体网关（Agent Gateway）策略控制着你的智能体可以与合作伙伴智能体共享哪些数据、可以基于合作伙伴的响应采取哪些行动，以及可以请求哪些信息。合作伙伴智能体在其自身的安全模型下运行，该模型已通过 Google Cloud 的集成要求验证。
设想一个多智能体支付工作流，其中你的智能体需要与外部商户智能体和金融服务智能体进行协调。各方在通过 A2A 协作处理共享交易的同时，都保持自己的治理体系。你的智能体无需了解商户系统内部是如何运作的。它通过 A2A 协议进行通信，双方独立执行各自的安全边界。
同样的原则适用于任何跨组织集成。你的智能体无需了解 Salesforce 的数据模型或 ServiceNow 的内部架构。它只需委托给智能体画廊中的合作伙伴智能体，该智能体由合作伙伴团队维护、经 Google Cloud 验证，并受你的智能体网关策略治理。当合作伙伴更新其平台时，他们的智能体也随之更新。你的系统无需任何改动即可受益。
## 模式五：环境事件网格（Ambient Event Mesh）
最后一种模式将 A2A 与事件驱动架构相结合，创建一个由环境智能体组成的网格，这些智能体在后台持续对事件做出反应，而无需等待用户请求。
Gemini Enterprise 智能体平台中的批处理与事件驱动智能体可直接连接 BigQuery 表和 Pub/Sub 流。将此与 A2A 相结合，你就创建了一个监听事件、处理事件并按需委托给专业智能体的智能体网络——全部在后台持续运行。
网格中的每个智能体都独立运行。当事件到达时，接收方智能体会处理它，并决定是本地处理、通过 A2A 委托给专业智能体，还是通过任务控制中心（Mission Control）升级给人工处理。网格是自组织的——添加一个新的专业智能体（例如欺诈检测专家）只需在智能体注册表中注册，并更新相关环境智能体中的路由逻辑。
这一模式对于事件量巨大的组织尤为强大。每天处理数百万次上传的内容平台、处理数百万笔交易的金融机构、处理数百万次客户互动的电商平台——都可以使用环境智能体网格来处理那些需要智能处理的长尾事件。
治理在这里至关重要。网格中的每个智能体都通过智能体身份拥有自己的身份标识。每次工具访问都由智能体网关治理。每次交互都通过智能体可观测性进行追踪。这个网格不是黑箱——它是一个完全可观测、受治理的专业智能体网络。
## 选择合适的模式
在构建多智能体系统时，请根据智能体之间的关系选择你的集成模式：
立即开始
互操作性技术栈现已可用：
- A2A 协议：ADK 支持 Python、TypeScript、Go 和 Java
- MCP：ADK 原生支持，GCP 数据库的托管支持
- 智能体画廊：Gemini Enterprise 中 100 多个经过验证的合作伙伴智能体
- 原生生态集成：面向企业系统的即插即用连接器
A2A + MCP 实验课程（codelab）演示了如何构建一个具有跨语言协作能力的多智能体系统：codelabs.developers.google.com/instavibe-adk-multi-agents
包含多智能体模式的 ADK 示例：github.com/google/adk-samples
构建孤立智能体的公司将在十二个月后被迫重构。而构建可互操作智能体生态系统的公司则会每天都在积累复利优势。
开始构建：adk.dev | https://cloud.google.com/products/gemini-enterprise-agent-platform
构建可互操作的智能体：Google for Startups AI Agents 挑战赛
孤立智能体是死路一条。加入我们为期 6 周的全球初创企业挑战赛，在 Gemini Enterprise 智能体平台上构建、优化或重构 AI 智能体。你将获得 500 美元云额度、完整的平台访问权限，并有机会争夺 9 万美元的奖金池。
立即报名，开始构建！
## 相关链接
- [Google Cloud Tech](https://x.com/GoogleCloudTech)
- [@GoogleCloudTech](https://x.com/GoogleCloudTech)
- [6.3K](https://x.com/GoogleCloudTech/status/2047567704807346675/analytics)
- [@addyosmani](https://x.com/@addyosmani)
- [@Saboo_Shubham_](https://x.com/@Saboo_Shubham_)
- [协调器-分发器架构](https://developers.googleblog.com/developers-guide-to-multi-agent-patterns-in-adk/)
- [ADK 集成生态系统](https://developers.googleblog.com/supercharge-your-ai-agents-adk-integrations-ecosystem/)
- [MCP Toolbox for Databases](https://adk.dev/integrations/mcp-toolbox-for-databases/)
- [codelabs.developers.google.com/instavibe-adk-multi-agents](https://codelabs.developers.google.com/instavibe-adk-multi-agents)
- [github.com/google/adk-samples](http://github.com/google/adk-samples)
- [adk.dev](https://adk.dev/)
- [https://cloud.google.com/products/gemini-enterprise-agent-platform](https://cloud.google.com/products/gemini-enterprise-agent-platform)
- [立即注册开始构建](https://devpost.team/hackathon_guest_invites/4fb181b4-2722-415d-a442-285a57dcaba5?utm_source=twitter&utm_medium=social&utm_campaign=google-for-startups-ai-agents-challenge&utm_content=twitter-post)
- [升级至 Premium](https://x.com/i/premium_sign_up)
- [2026年4月24日 下午2:45](https://x.com/GoogleCloudTech/status/2047567704807346675)
- [6,339 次浏览](https://x.com/GoogleCloudTech/status/2047567704807346675/analytics)
- [查看引用](https://x.com/GoogleCloudTech/status/2047567704807346675/quotes)
---
*导出时间: 2026/4/24 17:32:34*
