- ARD Release: Google has released Agentic Resource Discovery as an open way to publish and verify AI capabilities.
- Protocol Push: Google and Microsoft are backing a related backend-protocol effort against Anthropic and OpenAI influence.
- Stack Split: In 2025 and 2024 standards work, Agent2Agent coordinated agents while Model Context Protocol connected tools and data.
- Adoption Gate: Azure docs, partner catalogs, and registry support will determine whether ARD becomes an enterprise route.
Google has put its Agentic Resource Discovery (ARD) into public view on June 17, giving enterprise AI agents a specification for finding and verifying approved capabilities online. Agentic Resource Discovery lets organizations publish tools, skills, agent endpoints, and related metadata where trusted registries can discover them.
Google and Microsoft are also to be aligning around a broader AI backend-software protocol effort aimed at countering Anthropic and OpenAI influence. Against that reported alignment, Microsoft has not publicly confirmed an ARD-specific role; the public ARD specification sits with Google, while the Google-Microsoft lane remains a hedged competitive claim.
Enterprise customers now face a practical standards question. AI agents need approved routes to discover business tools, verify metadata, and connect to systems without turning every vendor integration into a one-off trust decision.
How ARD Turns Agent Discovery Into Infrastructure
Agentic Resource Discovery is built for software agents that need trusted directories for what they can use. ARD uses domain-hosted catalogs and registries so agents can discover MCP servers, A2A agents, and OpenAPI tools under an organization’s own web domain.
Within that infrastructure layer, a company domain can expose a catalog that names capabilities, sets interaction metadata, and gives registries a crawlable source for agent discovery. Google’s Gemini Enterprise Agent Platform Agent Registry is planned to support ARD for hosting and discovering capabilities.
Governance controls such as namespaced URNs, egress policies, and pinned specifications can control identity, network access, and version consistency before an agent connects. With that, security teams gain a new policy criterion before agents can reach external tools or internal business systems.
Multi-vendor agent deployments make that catalog layer commercially important. If ARD-style catalogs spread, platform owners can influence which agent ecosystems become default routes for business software without forcing every vendor to invent a separate integration path.
How A2A, MCP, and Rival Protocols Split the Stack
ARD addresses a different layer from agent-to-agent communication or tool and data connections. Google introduced the Agent2Agent protocol in 2025 so agents could communicate, securely exchange information, and coordinate actions across enterprise platforms.
Anthropic’s Model Context Protocol, introduced in 2024, connects AI assistants to systems where tools and business data live. Agent2Agent (A2A) is about coordination between agents, while Model Context Protocol (MCP) is about connecting an agent to the data and tools it needs.
ARD then adds a discovery and verification step before either kind of connection becomes useful. An agent has to know which capability exists, where it is published, whether the metadata is current, and whether the organization allows the connection.
OpenAI’s earlier adoption of Anthropic’s MCP shows why rival labs have a stake in protocol layers that can become enterprise defaults. Google Cloud’s deployment options for A2A agents added another route for customers that want agent coordination on cloud infrastructure.
The later established Agentic AI Foundation put Google, Microsoft, OpenAI, and Anthropic under Linux Foundation standards governance. Neutral control of agent plumbing is part of the enterprise adoption argument.
In June 2025, A2A moved under Linux Foundation governance after Google created it for secure agent collaboration. Microsoft backed that standards posture at the time by tying open interoperability to enterprise-grade controls for responsible deployment at scale.
Enterprise deployments often split into separate protocol layers: MCP for vertical tool integration and A2A or Agent Communication Protocol (ACP) for horizontal agent coordination. Agent Network Protocol adds an early decentralized alternative based on W3C decentralized identifiers.
Already more than hundres production-ready agentic AI tools now span 11 agent-tool categories. For enterprise AI teams, discovery, communication, and tool access can become separate architecture choices rather than one bundled model feature.
What Has to Be Proven Next
For Google and Microsoft’s effort, Salesforce, Snowflake, and ServiceNow matter because enterprise agent systems have to work inside business applications broadly.
ServiceNow already has a three-year OpenAI partnership for autonomous agents inside enterprise workflows, while Anthropic’s AWS work on an AWS AI agent marketplace shows how cloud and application vendors are competing over agent distribution.
A2A Release v1.0.1 arrived on May 28, 2026, giving developers a versioned protocol to test. ARD will need comparable evidence from Azure docs, registries, platforms, or partner catalogs before the reported alignment becomes a public enterprise adoption path.
