A coalition of major technology companies, including Google, Microsoft, GoDaddy, Hugging Face, NVIDIA, Salesforce, ServiceNow, Databricks, Snowflake, GitHub, and Cisco, has announced a new open standard called Agentic Resource Discovery (ARD). This initiative aims to solve a critical bottleneck in the development of autonomous AI agents: the inability to easily discover and connect to tools, skills, and other agents across different platforms and organizations.

AI agents are increasingly being deployed to perform complex tasks, but their effectiveness is often limited by the resources they can access. Without a standardized way to discover new capabilities, agents are confined to preconfigured tools, much like early internet users before search engines existed. ARD fills this gap by providing a unified discovery layer that allows agents to query and find relevant resources at runtime.

The Discovery Gap

In 2024, Anthropic introduced the Model Context Protocol (MCP), which standardized how AI systems communicate with servers to share data. While MCP solved the connection problem, it left a critical issue: how do agents find the available servers and tools in the first place? This is analogous to having an app store without a search function or catalog. ARD is designed to be that catalog, enabling agents to dynamically discover and evaluate resources.

Ramanathan Guha, Technical Fellow at Microsoft, explained that AI is currently limited by its wiring. Agents can only use what they have been explicitly connected to. Everything else might as well not exist. This limitation prevents agents from adapting to new situations and leveraging the vast ecosystem of tools and services available on the web. ARD removes this barrier by creating a decentralized registry where agents can search for capabilities, verify their source, and connect to them.

How ARD Works

The ARD specification defines two main components: catalogs and registries. Catalogs are JSON files (named ai-catalog.json) hosted on an organization's domain. They list the resources—tools, skills, agents—that the organization makes available. Registries act as search engines for the agentic web. They crawl catalogs, index their contents, and return matching capabilities with metadata to verify the publisher before connection.

Domain ownership serves as the cryptographic foundation for identity and trust. Because a catalog is hosted on a trusted domain like microsoft.com or googledotcom, agents can have some confidence that the resources listed have been vetted by the domain owner. This hierarchy mirrors the Domain Name System (DNS), giving ARD an architectural property closer to DNS than to ordinary web search. Microsoft's Guha emphasized that this design allows ARD to scale securely and efficiently.

Security Considerations

While the domain-based trust model reduces some risks, it also creates new attack surfaces. Attackers may target domains, deployment pipelines, or catalog files to inject malicious resources. ARD is designed to sit before invocation, helping an AI client choose a capability before connecting. But once a compromised catalog is accepted, the agent could be tricked into using a harmful tool.

Google and Microsoft have acknowledged these concerns and recommend additional enterprise controls such as Agent Identity, trust manifests, egress policies, and pinned tools. However, the open-web model remains domain-anchored, meaning that if the domain, DNS, server, repository, or deployment path is compromised, the catalog becomes a high-leverage target. ARD improves discovery and verification but does not eliminate the need for robust security practices, including allowlists, code review, signing, monitoring, and policy enforcement.

Reference Implementations

Several vendors have already integrated ARD into their platforms. GitHub launched Agent Finder, which allows GitHub Copilot to discover and call MCP servers, skills, tools, and agents at runtime from public or private registries. Hugging Face introduced a Discover Tool that offers semantic search to thousands of Skills and MCP Servers. Google plans to support ARD through its Agent Registry in the Gemini Enterprise Agent Platform in the coming months.

These implementations demonstrate the potential of ARD to unlock new levels of interoperability. For example, an enterprise agent could search for a specific data processing skill, verify its authenticity via domain metadata, and use it to complete a task—all without human intervention. This reduces the burden on developers to manually configure each connection and allows agents to evolve dynamically.

The open specification is licensed under Apache 2.0 and is built on the AI Catalog data model from a Linux Foundation working group. The full specification is available at AgenticResourceDiscovery.org, and a GitHub registry hosts the spec for community contributions. The coalition invites other organizations to participate, emphasizing that the agent ecosystem works best when decentralized and open.

Notably absent from the announcement are OpenAI and Anthropic, two of the leading AI companies. Their absence raises questions about whether they will adopt ARD or develop their own discovery standards. The fragmentation of discovery protocols could undermine the goal of universal interoperability. However, the involvement of major cloud platforms and enterprise software vendors suggests strong momentum.

ARD represents a significant step forward for AI agent ecosystems. By providing a standardized way to discover and verify resources, it promises to make agents more capable, adaptable, and collaborative. Yet, as with any open system, security remains a paramount concern. Organizations adopting ARD must implement layered defenses to protect against exploitation. The coming months will reveal whether this new standard can deliver on its promise without creating new vulnerabilities.

Source: ZDNET News