Agentic Resource Discovery: Draft Specification Lets Agents Search for Tools at Runtime

Currently, agent capabilities follow an install-first, use-later model where developers hardcode server URLs or connect services via plugins. [1] Alternatively, developers must dump every single available tool description into a model's context window, a strategy which is ultimately limited by the system's strict context budget. [1]

The Agentic Resource Discovery specification is a draft, open standard developed with industry participation to define how tools and agents are cataloged and searched. [1] ARD moves the selection process outside the LLM by indexing capabilities with signals like publisher identity, representative queries, and compliance attestations through a REST endpoint. [1] Agents can search for capabilities in natural language and invoke the returns without pre-configuring them. [1]

ARD shifts the paradigm from static configurations—where developers manually integrate specific tools—to dynamic, intent-based discovery. While existing protocols like MCP give agents a standard way to execute tools, ARD sits in front of them to handle the search and cataloging step. This could alleviate context window exhaustion by querying only relevant capabilities when they are actually needed, rather than loading them all at initialization. What the sources don't address: How the specification intends to handle the security and authentication risks associated with dynamically discovering and executing unseen third-party tools at runtime.