Eleven Vendors Shipped an Agent Discovery Standard Without OpenAI

On June 17, eleven companies published Agentic Resource Discovery — ARD for short — an open specification for how AI agents find, verify, and connect to tools, skills, and other agents across corporate software stacks. The contributor list reads like a roll call of enterprise incumbency: Google, Microsoft, Salesforce, Cisco, Databricks, GitHub, GoDaddy, Hugging Face, NVIDIA, ServiceNow, and Snowflake. OpenAI and Anthropic are nowhere on it. That absence is not an oversight. It is the story. ARD is not a model benchmark or a safety framework. It is infrastructure for a bet that the winning interface for workplace AI will not be a chat window owned by a frontier lab, but the software suites enterprises already pay for — with agents that discover capabilities dynamically instead of waiting for a human to wire every connection by hand.

The technical problem ARD targets is mundane and expensive. Today, every agent must be pre-configured with every MCP server, API, skill, and sub-agent it might need. A power user might wire dozens of connectors into GitHub Copilot; most installations expose only a handful. Meanwhile, the ecosystem already contains hundreds of thousands of callable resources scattered across repositories, vendor catalogs, and internal codebases. Microsoft's Ramanathan Guha, a Technical Fellow who co-authored the specification, put the gap in plain terms: "AI is only as capable as its wiring allows." ARD moves selection outside the model. Publishers host an ai-catalog.json manifest at a well-known path on their domain; registries crawl and index those catalogs; agents query a REST search endpoint in natural language and receive ranked matches with publisher identity, representative queries, and trust metadata. Discovery is separated from execution — once a capability is chosen, the agent connects through whatever protocol the tool uses: MCP, Google's Agent2Agent, a plain API, or something else.

The coalition shipped working implementations the same week it published the draft. GitHub launched agent finder, letting Copilot discover MCP servers, skills, tools, and agents from a curated catalog or a private registry at runtime instead of pre-loading everything into the context window. Hugging Face released its Discover Tool as a reference registry, indexing thousands of Spaces, skills, and MCP servers on the Hub. Cisco tied ARD to its AGNTCY Agent Directory under the Linux Foundation. Google said its Gemini Enterprise Agent Registry will add native ARD support in the coming months — support that is not live yet. The specification is version 0.9, licensed under Apache 2.0, and built on the ai-catalog data model maintained by a Linux Foundation working group. Contributors are soliciting changes through the project's GitHub repository. This is early infrastructure, not finished plumbing. But the timing is deliberate: the same week OpenAI filed for its IPO, Anthropic faced export-control enforcement on frontier models, and 42 state attorneys general subpoenaed OpenAI over model behavior, the enterprise stack declared how agents will route themselves.

The strategic split is sharper than the polite language suggests. OpenAI and Anthropic have built enterprise businesses around conversational interfaces — ChatGPT Enterprise, Claude for Work, custom connectors, and the assumption that the model is the front door. ARD assumes the opposite: the front door is whatever application the employee already has open, and the model is a runtime component that discovers what it needs. Crypto Briefing's June 20 analysis framed the coalition as drawing "a line in the sand" — a vision where agents "operate autonomously within existing software ecosystems, discovering and using tools without a conversational interface as the bottleneck." That is not a technical nuance. It is a distribution fight. A world where Copilot, Gemini Enterprise, and Salesforce agents automatically locate and invoke the right corporate tools is a world where Google Workspace, Microsoft 365, and Salesforce CRM remain the billing relationship. Guha made the historical analogy explicit: early web users visited only bookmarked sites until search engines "lit the web up." ARD, he wrote, "does the same for the agentic resource ecosystem." He also compared the architecture to DNS — local control, upstream sources, federated resolution — rather than to a single global search index.

The specification's design choices reinforce incumbent advantage without requiring monopoly ownership. ARD is federated: each organization publishes its own catalog, each registry controls its answer set, and enterprises can merge private resources with vetted vendor and public capabilities. Authentication and authorization remain delegated to the underlying protocols; ARD handles discovery and verification, not permissioning. That is why the verification step matters as much as the search step. A Hugging Face community comment on the launch post put it bluntly: "Discovery without verification just industrialises trusting strangers." The spec supports cryptographic trust manifests, domain-anchored URNs, and attestations — signals richer than dumping tool descriptions into a context window and hoping the model picks correctly. For publishers of callable software — APIs, MCP servers, internal agents — ARD is immediately actionable: host a manifest, get indexed, become discoverable. For content sites and storefronts, as several analysts have noted, there is nothing to optimize today. The standard targets the machine-actionable layer of the web, not the human-readable one.

OpenAI and Anthropic are not locked out forever. ARD's authors say it complements Anthropic's Model Context Protocol rather than replacing it; MCP handles invocation, ARD handles discovery. Any lab could publish an ai-catalog.json and ship a registry. But standards wars are won by who shows up first with distribution, and the first mover list is telling. NVIDIA is on it — the chip layer wants agents to find tools, not just models. Snowflake and Databricks are on it — the data layer does too. Google is on it — and just lost Noam Shazeer, co-inventor of the transformer architecture, to OpenAI days before the latter filed for its own public offering. The enterprise coalition is not waiting for frontier labs to agree. It is building the directory while the labs fight over model weights, export controls, and IPO disclosures. Anthropic invented MCP and made tool use a category; Google and Microsoft are now building the phone book that decides which tools get called.

For buyers and builders, ARD changes the procurement calculus. Agent governance has focused on permissions, logging, and human approval — necessary but insufficient if agents cannot safely discover what they are permitted to use. ARD introduces a new layer: catalog hygiene. Which tools are published? Who attested them? Which registry does the organization trust? Does the agent search a public index, a vendor catalog, or an internal one? Estonia's push for government-backed agent identities, covered elsewhere this week, addresses who the agent is. ARD addresses what the agent can find. Together they sketch an agentic internet that looks less like a chatbot and more like regulated infrastructure. The v0.9 draft leaves plenty unresolved — federation modes are still maturing, Google's Agent Registry integration is months away, and no one has proven registries can crawl at web scale yet. But the direction is clear: dynamic discovery replaces static installation, and the companies that control discovery inherit leverage over every agent that depends on it.

Convina's view: ARD is the most important enterprise AI story of the week that is not about models, money, or Washington — and precisely because it is about all three indirectly. The frontier labs want to own the interface. The enterprise stack wants to own the routing layer. ARD is the routing layer, published as an open standard but backed by the vendors whose subscriptions agents will traverse. OpenAI's IPO filing asks public markets to bet on ChatGPT as the front door to work. Anthropic's export-control fight asks Washington to treat its models as strategic infrastructure. ARD asks something simpler and harder to IPO around: publish your tools, let agents search, and may the best catalog win. The labs are not on the initial supporter list because joining would mean conceding that the chatbot may not be the product — the directory is. Every enterprise building on frontier models should assume their agents will soon search someone else's registry. The question is whether your tools are in it.