AI News
官方更新 AWS ML Blog Mon, 24 Aug 2026 正文 ok 1014 words

Agentic Resource Discovery（ARD）：面向 Agent 发现的开放规范

英文标题：Agentic Resource Discovery (ARD): An open specification for agent discovery

AWS Agent Registry 为企业提供集中式、可搜索的目录，用于管理 Agent、MCP 服务器、工具、Agent 技能及自定义资源，并通过审批、授权和混合搜索支持治理。随着 Agent 和工具分布在多云、本地部署与 SaaS 环境中，统一发现面临格式、命名和元数据不兼容的问题。开放规范 Agentic Resource Discovery（ARD）通过通用资源描述格式和发现协议，旨在让不同环境的注册表实现联邦式互操作。AWS 认为，ARD 可与 AWS Agent Registry 互补，在保留本地访问控制的同时支持跨环境乃至跨组织的资源发现。
打开原文相关 2.7 · 机会 4
media1.3 MB media101.1 KB media124.1 KB media345.0 KB media141.2 KB media346.5 KB media19.3 KB media17.9 KB media25.0 KB media114.9 KB

中文

AWS Agent Registry 与 Agentic Resource Discovery（ARD）规范如何实现 Agent 的跨环境发现

随着组织不断扩大对人工智能（AI）Agent 和工具的使用规模，找到合适的资源正成为难题。团队构建 Model Context Protocol（MCP）服务器、部署 Agent 并创建专用工具，但如果没有统一目录，这些资源就会形成孤岛。开发者需要手动定位资源、审查资源、连接资源并维护这种连接。更糟的是，为一个 AI 客户端配置 Agent，并不会让它自动可供另一个客户端使用。

当一个团队仅需连接少量工具时，这种方式尚可管理。但面对如今分散在公共注册表和企业私有环境中的大量 Agent、MCP 服务器、技能和 API，这种方式无法扩展。

AWS Agent Registry：集中式、可搜索的目录

AWS Agent Registry 为组织提供集中式目录，用于管理 Agent、MCP 服务器、工具、Agent 技能和自定义资源。它围绕两个核心概念构建：

  • 注册表（Registries）。 注册表是在 AWS 账户中创建的目录，拥有独立的授权配置和审批设置。你可以运行一个覆盖全组织的注册表，也可以按资源类型、阶段或团队分别建立注册表。通过跨账户共享，一个注册表可以服务于整个 AWS Organization。
  • 注册表记录（Registry Records）。 一条记录代表一个独立资源，包含用于说明该资源是什么、能做什么以及如何访问它的元数据。

工作流程很直接

  • 创建注册表： 管理员创建注册表、配置审批设置，并使用 AWS Identity and Access Management（IAM）或企业身份提供商签发的 JSON Web Token（JWT）设置授权。
  • 发布记录： 发布者将其 MCP 服务器、Agent 或工具描述为记录，并提交审批。
  • 整理记录并审批： 管理员审核待处理记录，批准或拒绝记录，并废弃不再使用的记录。
  • 发现已批准的资源： 使用者无论是人类用户还是 AI Agent，都可以在注册表中搜索所需资源。

企业级能力

  • 治理： 审批工作流可确保只有符合安全、合规和质量标准的记录能够被发现。管理员可随时将记录从发现结果中移除。
  • 混合搜索： 结合语义理解与关键词匹配，因此无论是自然语言查询还是精确名称查询，都能返回相关结果。
  • MCP 原生访问： 注册表可通过远程 MCP 端点访问，因此任何兼容 MCP 的客户端都可以直接搜索和使用它。
  • 灵活授权： 可通过 IAM 凭证或企业身份提供商签发的 JWT 控制访问权限。

多环境挑战

AWS Agent Registry 解决了 AWS 环境内部的资源发现问题。但大多数企业并非只在一个环境中运行。Agent 和工具通常部署在多个云平台、本地基础设施、SaaS 平台和企业应用中，每个环境都有自己的注册表、命名约定和元数据模式。

当每个环境都使用不同格式来描述 Agent 资源时，要将它们整合起来，就需要为每一对需要互操作的注册表构建定制连接器。共享规范能够改变这一局面：如果每个注册表都以相同格式描述资源，并通过通用协议提供发现能力，那么发布者只需描述一次，使用者便能在任何地方发现资源。

Agentic Resource Discovery（ARD）登场

ARD 是一项开放标准，而非某个产品或单一注册表。它采用 Apache License 2.0 许可，可在 agenticresourcediscovery.org 和 GitHub 获取。AWS 在该规范开发期间提供了反馈。

可以将 ARD 理解为支持注册表之间联邦的机制，类似于 Domain Name System（DNS）如何实现跨网络名称解析。组织可以在不同环境中部署 Agent，而每个环境的目录都会通过端点，使用通用协议呈现这些资源。为了实现组合式发现，任何注册表都可以基于对这一共享通用协议的理解，跨多个目录建立索引。因此，本地注册表可以通过 ARD 实现联邦，而无需双边协议或专有连接器。

ARD 如何补充 AWS Agent Registry

我们认为，ARD 是对 AWS Agent Registry 模式的自然补充：

  • 无需迁移即可联邦： 在云端、本地环境和 SaaS 中分布 Agent 基础设施的组织，可以通过一致的格式暴露这些资源。我们预计，ARD 将在保持本地控制权的同时实现跨环境发现。
  • 全局发现，本地控制： ARD 的设计反映了 AWS 客户所期待的控制模式：发布目录的组织可控制目录中包含什么、谁能够看到它，以及何时撤销访问权限。我们预计，AWS Agent Registry 现有的访问控制仍将处于执行点，ARD 则充当互操作层。
  • 实现公共发现： 通过 ARD 这一共享协议，任何组织都可以在自己的域名上发布目录，使其能够被任何兼容 ARD 的客户端发现。我们预计，ARD 将为 Agent Registry 客户开辟跨组织的资源发现路径。

了解更多

这只是一个开始。随着 AWS Agent Registry 以及我们对开放发现标准的支持持续演进，我们期待获得你的反馈。

作者简介

Jeffrey Damick

Jeffrey 是 Amazon Web Services 的首席软件工程师，负责 Amazon Route 53 以及大规模 DNS 和网络技术。他近期的工作聚焦于 DNS 与人工智能的交叉领域，探索成熟的互联网基础设施如何支持 AI Agent 的发现与通信。这项工作包括演进基础互联网协议和标准，以实现对 Agent 及其能力的可扩展、开放式发现。

Bhargav Talluri

Bhargav 是 AWS 的高级产品管理经理，负责领导 Amazon Route 53、AWS Agent Registry 中的 Agent 身份与发现，以及 Amazon Domains Registrar 业务的产品路线图。除管理权威 DNS、递归 DNS 和安全产品组合外，他还专注于建设相关基础设施，使 AI Agent 能够在开放网络中获得稳定、可验证的身份，覆盖从基于 DNS 的发现到通过开放标准实现跨生态系统联邦。

Anubhav Mangal

Anubhav 是 Amazon Web Services 的首席产品经理，任职于 Amazon Bedrock AgentCore，负责 Agent 资源的发现与治理以及 AWS Agent Registry。他此前曾担任 AWS Marketplace 高级产品管理经理。他的职业生涯横跨印度和美国的技术与咨询领域，拥有计算机科学硕士学位和工商管理硕士学位。工作之余，他喜欢阅读小说、徒步和烘焙，也关注足球和一级方程式赛车。

English / 原文

How AWS Agent Registry and the Agentic Resource Discovery (ARD) specification enable cross-environment discovery for your agents

As organizations scale their use of artificial intelligence (AI) agents and tools, finding the right resource becomes the hard part. Teams build Model Context Protocol (MCP) servers, deploy agents, and create specialized tools, but without a central catalog, those resources stay siloed. Developers locate a resource, vet it, connect it, and maintain that connection manually. Worse, configuring an agent for one AI client doesn’t make it available to another.

This was manageable when a team wired up a handful of tools. It doesn’t scale to a growing number of agents, MCP servers, skills, and APIs now spread across public registries and private enterprise estates.

AWS Agent Registry: A centralized, searchable catalog

AWS Agent Registry gives your organization a centralized catalog for agents, MCP servers, tools, agent skills, and custom resources. It’s built around two core concepts:

  • Registries. A registry is a catalog you create in your AWS account with its own authorization configuration and approval settings. You can run a single org-wide registry or separate registries by resource type, stage, or team. With cross-account sharing, a registry can serve your entire AWS Organization.
  • Registry Records. A record represents an individual resource, capturing metadata that describes what it is, what it does, and how to reach it.

The workflow is straightforward

  • Create a registry : An administrator creates a registry, configures approval settings, and sets up authorization using either AWS Identity and Access Management (IAM) or a JSON Web Token (JWT) from your corporate identity provider.
  • Publish records : A publisher describes their MCP servers, agents, or tools as records and submits them for approval.
  • Curate records and approve : A curator reviews pending records, approves or rejects them, and deprecates records no longer in use.
  • Discover approved resources: Consumers, whether human users or AI agents, search the registry for the resources they need.

What makes it enterprise-ready

  • Curation: An approval workflow ensures only records that meet your security, compliance, and quality bar are discoverable. Administrators can remove a record from discovery at any time.
  • Hybrid search: Semantic understanding combined with keyword matching, so both natural-language queries and exact name lookups return relevant results.
  • MCP-native access: The registry is available at a remote MCP endpoint, so any MCP-compatible client can search and use it directly.
  • Flexible authorization: Control access with IAM credentials or JWTs from your corporate identity provider.

The multi-environment challenge

AWS Agent Registry solves discovery within your AWS environment. But most enterprises don’t operate in one place. Agents and tools are deployed across multiple clouds, on-premises infrastructure, SaaS platforms, and enterprise applications, each with its own registry, naming convention, and metadata schema.

When each environment uses its own format for describing agentic resources, bringing it all together requires bespoke connectors for every pair of registries that need to interoperate. A shared specification changes that equation: if every registry describes resources in the same format and exposes discovery through a common protocol, publishers describe once and consumers discover everywhere.

Enter: Agentic Resource Discovery (ARD)

ARD is an open standard, not a product or a single registry. It’s available under the Apache License 2.0 at agenticresourcediscovery.org and on GitHub. AWS contributed feedback during the spec’s development.

Think of ARD as enabling federation across registries analogous to how the Domain Name System (DNS) enables name resolution across networks. An organization can deploy agents across environments, and each environment’s catalog surfaces these resources in a common protocol, behind an endpoint. For combined discovery, any registry can index across them using an understanding of the shared common protocol. Thus, local registries can federate through ARD without requiring bilateral agreements or proprietary connectors.

How ARD can complement AWS Agent Registry

We see ARD as a natural complement to the AWS Agent Registry model:

  • Federate without migrating: Organizations with agentic infrastructure spread across clouds, on-premises, and SaaS could expose those resources in one consistent format. We expect ARD to enable cross-environment discovery while keeping control local.
  • Discover globally, control locally: ARD’s design mirrors the control model AWS customers expect: the organization that publishes a catalog controls what’s in it, who can see it, and when to revoke access. We expect the existing access controls of AWS Agent Registry to remain at the enforcement point, with ARD serving as the interoperability layer.
  • Enable public discovery: With ARD as a shared protocol, any organization can publish a catalog on its own domain, making it discoverable by any ARD-compatible client. We expect ARD to open cross-organizational discovery paths for Agent Registry customers.

Learn more

This is only the beginning. We’d love your feedback as AWS Agent Registry and our support for open discovery standards evolve.

About the authors

Jeffrey Damick

Jeffrey is a Principal Software Engineer at Amazon Web Services, where he works on Amazon Route 53 and large-scale DNS and networking technologies. His recent work focuses on the intersection of DNS and artificial intelligence, exploring how established Internet infrastructure can support the discovery and communication of AI agents. This work includes evolving foundational Internet protocols and standards to enable scalable, open discovery of agents and their capabilities.

Bhargav Talluri

Bhargav is a Senior Manager, Product Management at AWS, where he leads the product roadmap for Amazon Route 53, Agent identity and discovery in AWS Agent Registry, and the Amazon Domains Registrar business. In addition to managing the product portfolio for Authoritative & Recursive DNS and Security products, he focuses on building the infrastructure that gives AI agents stable, verifiable identities on the open web from DNS-based discovery to cross-ecosystem federation via open standards.

Anubhav Mangal

Anubhav is a Principal Product Manager at Amazon Web Services with Amazon Bedrock AgentCore, where he owns discovery of governance of agentic resources and AWS Agent Registry. He was formerly Sr. Manager, Product Management at AWS Marketplace. He has spent his career in technology and consulting across India and the US. He has a Masters in Computer Science and a Masters in Business Administration. Outside of work, he enjoys reading fiction, hiking, baking, and follows soccer and Formula One.