AIエージェントが自分でツールを探す時代へ！新仕様「ARD (Agentic Resource Discovery)」超入門The proposed ARD (Agentic Resource Discovery) spec extends MCP so AI agents can…

A newly proposed specification called ARD, short for Agentic Resource Discovery, aims to extend the Model Context Protocol (MCP) so that AI agents can search for and connect to tools on their own at runtime. The idea matters because it targets one of the more practical bottlenecks in building agent systems today: the need to know, in advance, exactly which tools and data sources an agent will use. If adopted, ARD could shift agent design away from static, hardcoded configurations toward more dynamic and flexible architectures.

To understand the proposal, it helps to review MCP itself. Introduced by Anthropic in late 2024 and since embraced by a growing number of vendors, MCP is an open protocol that standardizes how AI applications connect to external tools, data, and prompts. In the typical setup, an MCP client, often embedded in a chat app or IDE assistant, connects to one or more MCP servers that each expose a defined set of capabilities. This standardization has made it far easier to plug models into file systems, databases, APIs, and third-party services without writing bespoke integration code for every combination.

The limitation ARD appears to address is discovery. In current MCP deployments, the servers an agent can reach are usually declared ahead of time in a configuration file or application settings. That works well when an agent needs a handful of known tools, but it becomes cumbersome as ecosystems grow into hundreds or thousands of available servers. Developers must anticipate which capabilities will be needed, maintain connection details manually, and update configurations whenever the toolset changes. ARD is positioned as a way to let an agent query for relevant resources on demand, much as a person might search a catalog rather than memorizing every entry.

Based on its description, ARD would introduce a discovery layer on top of MCP in which an agent can issue a search, describe the capability it is looking for, and receive candidate tools or resources it can then invoke. This would reduce the amount of information that has to be baked into an agent at build time and allow the same agent to adapt as new tools appear. The approach is conceptually similar to service discovery in distributed systems and microservices, where components locate one another at runtime through a registry rather than through fixed addresses. The specific mechanics, including how results are ranked, how trust and permissions are handled, and how an agent verifies that a discovered tool does what it claims, are the kinds of details that will determine how useful the specification proves to be in practice.

ARD does not exist in isolation. The broader industry has been moving toward interoperability standards for agents and tools, including Google's Agent2Agent (A2A) protocol for agent-to-agent communication and various emerging registries and marketplaces for MCP servers. Anthropic and others have also discussed the need for discovery and governance as MCP adoption scales. Concepts such as tool registries, capability manifests, and semantic search over tool descriptions are recurring themes across these efforts, and ARD appears to sit squarely within that trend rather than replacing MCP or competing frameworks.

Several open questions accompany any runtime discovery model. Allowing an agent to find and call tools it was not explicitly granted raises security and safety considerations, since an autonomously discovered tool could behave unexpectedly or expose sensitive data. Reliability and reproducibility can also become harder to guarantee when the set of available tools changes between runs. Proponents of dynamic discovery generally argue that these tradeoffs are manageable with appropriate authentication, sandboxing, and human oversight, but such safeguards would need to be specified clearly for ARD to be trusted in production environments.

For now, ARD is best understood as a proposal that reflects where agent tooling is heading rather than a finalized, widely implemented standard. Readers building on MCP may want to track how the specification develops, whether major clients and server authors adopt it, and how it aligns with related discovery and interoperability initiatives. If it gains traction, the ability for agents to locate their own tools could become a meaningful step toward more autonomous and maintainable systems.