In June 2026, Google and ten of the world’s largest technology companies published a draft specification called ARD, Agentic Resource Discovery, that defines how AI agents find the tools and capabilities they need at runtime. The backing coalition includes Microsoft, Cisco, Databricks, GitHub, GoDaddy, Hugging Face, NVIDIA, Salesforce, ServiceNow, and Snowflake. Together, these eleven organizations own the infrastructure, data platforms, developer tooling, and enterprise software layers that underpin most large organizations globally.
The signal beneath the announcement: Anthropic’s Model Context Protocol has become the de facto connection standard for AI agents over the past eighteen months. ARD is the industry’s answer to that dominance, and it is backed by players with enough combined market weight to make a competing standard stick.
The Problem ARD Solves
MCP defines how an AI model connects to a specific tool. What it does not define is how an agent finds the right tool in the first place.
Consider a realistic enterprise scenario: an agent is tasked with compiling a competitive briefing, scheduling a follow-up meeting, and logging a deal update in the CRM. That agent needs to discover, in real time, which MCP servers handle calendar access, which API manages the CRM, and which data connector reaches competitive intelligence sources, without a human pre-configuring each one.
Without a discovery layer, every enterprise agent deployment requires manual wiring. The agent builder has to know which tools exist, where they live, and how to connect them before writing a single line of agent logic. At the scale of dozens of internal tools across a mid-size enterprise, this becomes a configuration management problem that defeats the autonomous value proposition of agents entirely.
ARD is the pre-invocation layer. It answers “what is available?” before MCP answers “how do I connect?” Hugging Face described it as enabling “intent-based search that lets an agent find the right capability dynamically and reach a growing ecosystem of MCP tools, A2A agents, and other services without pre-configuring each one.”
How ARD Works
The specification has two components: catalogs and registries.
Catalogs: An organization publishes an ai-catalog.json file at /.well-known/ai-catalog.json on their domain. The file lists the AI capabilities the organization makes available: MCP servers, A2A agents, OpenAPI tools, or nested catalogs pointing to other organizations. Each entry includes capability metadata describing what the tool does in terms an agent can reason about, plus a trust manifest.
The trust manifest is the enterprise-grade element. Each catalog entry can carry SOC2, HIPAA, and GDPR compliance attestations, plus cryptographic identity via SPIFFE or DID. An agent querying the catalog can verify the publisher’s identity and compliance posture before invoking any tool, without a human approving each connection. This is the audit trail that enterprise compliance and security teams require before they allow autonomous agents to call external systems.
Registries: Registries are search engines for AI capabilities. They crawl published catalogs, index capabilities, and expose a search API that agents query at runtime. An agent searching for a calendar integration does not scan every domain on the internet; it queries a registry, receives a ranked list of verified options, and selects the one matching its task context and trust requirements. Google Cloud’s Agent Registry, embedded in the Gemini Enterprise Agent Platform, is the first production registry implementation aligned to the ARD specification.
The full discovery flow: organization publishes a catalog at the well-known path, registry crawls and indexes it, agent searches by intent, agent verifies the publisher, agent connects via MCP or A2A or a plain API.
What the Coalition Signals
The composition of the ARD working group is as important as the technical specification.
Anthropic’s MCP crossed 97 million monthly SDK downloads by February 2026 and was adopted by every major AI provider. The Linux Foundation formally incorporated MCP and A2A under the Agentic AI Foundation in December 2025, with OpenAI, Anthropic, Google, Microsoft, AWS, and Block as co-founders. That should have settled the standards race.
It did not. CryptoBriefing reported that ARD is framed explicitly as a counter to Anthropic and OpenAI in enterprise agent infrastructure, targeting the plumbing layer where the most durable competitive advantages are built. Salesforce, Snowflake, and ServiceNow collectively touch most of the world’s enterprise data and workflows. Google and Microsoft own the clouds those workflows run on. Their public alignment on a shared spec signals that the incumbents view the agent infrastructure layer as too strategically significant to cede to foundation model providers.
Search Engine Journal noted that ARD does not replace MCP or A2A: it sits above them. An enterprise can adopt ARD for discovery while still using MCP for every tool connection. The standards are additive, not competing at the protocol level. The competition is at the governance level: who defines the trust model for enterprise agent ecosystems.
Protocol Comparison: ARD, MCP, and A2A
| Layer | Standard | What It Does | Lead Backer |
|---|---|---|---|
| Discovery | ARD | Catalogs and searches agent capabilities before connection | Google, Microsoft |
| Connection | MCP | Connects AI to a specific tool or data source | Anthropic |
| Agent-to-Agent | A2A | Defines how agents delegate tasks to other agents | |
| Foundation | AAIF (Linux Foundation) | Governs MCP and A2A long-term | Cross-industry |
All four layers will coexist in production enterprise agent stacks. An agent finds a tool via ARD, connects via MCP, delegates sub-tasks via A2A, and operates under governance rules set by the AAIF-maintained specifications. The layers do not cancel each other.
Enterprise Implications
For teams building internal agent infrastructure, ARD changes the design question from “which tools does this agent know about?” to “which tools can any agent in our organization discover?”
An enterprise that publishes ARD-compatible catalogs for its internal systems, CRM, ERP, scheduling, document management, and data warehouse, enables any agent deployed across the organization to discover those capabilities at runtime without bespoke integration work per agent. The catalog becomes a single source of truth for what is available, verified, and compliant.
For procurement and security teams, the trust manifest embedded in ARD catalogs is the more immediate value. Rather than manually vetting each tool connection a new agent requires, security teams can define which compliance attestations are required and let agents self-select verified tools from a registry. This does not eliminate security review, but it structures the review at the catalog level rather than at each individual invocation.
For platform and SaaS vendors, the implication is direct: publish an ai-catalog.json or accept that enterprise agents will not discover your product unless a human manually configures the connection. By 2027, ARD compatibility will likely appear as a line item in enterprise procurement checklists for the same reason REST API documentation became standard: buyers expect machine-readable integration surfaces.
Where Things Stand Today
ARD is a v0.9 draft specification with near-zero measured adoption. A June 2026 census of 39 major domains, including all eleven working group members, found zero live ai-catalog.json files. The specification is available under an open license from the Agentic Resource Discovery organization.
The gap between announcement and adoption is expected. OpenAPI took years from its 2011 Swagger origins to become a procurement requirement. MCP took roughly six months from its November 2024 launch to 97 million monthly downloads. ARD’s timetable will depend on how quickly Google Cloud’s Agent Registry drives production adoption through Gemini Enterprise deployments and how quickly Salesforce, ServiceNow, and Snowflake surface ARD catalog requirements in their own agent platforms.
The enterprises best positioned are those that treat the announcement as a design constraint today: build internal tool catalogs in a format that can be serialized to ARD-compatible JSON when the tooling matures, rather than rebuilding catalog infrastructure from scratch when the standard reaches v1.0.
As the enterprise AI adoption gap narrows, the organizations that move from manual agent wiring to catalog-driven discovery will deploy agents faster and govern them more reliably. ARD is the infrastructure decision that determines which side of that gap an enterprise sits on in 2027.
The battle for the enterprise AI workspace is being fought at the product layer by Anthropic, OpenAI, and Google. ARD signals that a parallel battle is being fought at the infrastructure layer, and it has eleven of the most consequential players in enterprise technology aligned behind it.
For enterprise teams building or buying agent infrastructure right now: the product you pick for agent execution matters. The catalog you publish, or fail to publish, will determine whether autonomous agents can find your tools in the first place.