MCP was supposed to solve the agent tooling problem. It missed a step.
The New Stack Amanda Caswell
AWS is backing a new spec that helps AI agents find tools across catalogs. It tries to fix the bit MCP skips: discovery before connection.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Connecting an AI agent to one tool is easy enough. The mess starts when an organization has hundreds or thousands of resources spread across clouds, SaaS platforms, and internal systems. That is the gap Agentic Resource Discovery, or ARD, is trying to fill.
AWS put ARD in its August 31 Weekly Roundup after taking a closer technical look a week earlier, calling it “DNS, but for agents.” The comparison is useful, but only up to a point. ARD does not point an agent to one fixed place. It lets the agent search across registries and see what’s available before it tries to use anything.
That matters because MCP, which is now a common way for AI apps to connect to outside tools and data, assumes the client already knows which server it wants. ARD changes that by adding a discovery layer. The spec uses “agentic resource” for anything an AI client can connect to, from an MCP server to other outside capabilities, and an ARD-compatible service keeps track of what is available instead of forcing developers to wire up every connection in advance.
The project is not an AWS invention, despite the company’s enthusiasm. It was written by Junjie Bu of Google, R.V. Guha of Microsoft, and Shaun Smith of Hugging Face, and released under Apache 2.0. Engineers from Cisco, Databricks, GitHub, GoDaddy, Nvidia, Salesforce, ServiceNow, and Snowflake have also shaped it. AWS says its job so far has been feedback and experimentation with its own Agent Registry.
The current v0.91 proposal, dated August 26, uses JSON-LD and a REST interface. Its required POST /search endpoint looks up resources by task, while optional endpoints let clients browse what a discovery service knows about. Each service can also decide what it returns and which sources it trusts, which is why this is more federation than a single central directory. The appeal is obvious: an enterprise can keep private resources private and still search approved external catalogs when needed.
But the governance side is still loose. Board terms, membership, and even the question of whether ARD should eventually move to a neutral home like the W3C or an AI foundation are still open. And AWS is already exploring how ARD could fit with Agent Registry, including resources outside its own catalog. The longer companies wait, the more discovery systems they will build on their own, and the harder this gets.
My take — AI-written commentary, not fact-checked reporting
MCP got the buzz because it solved the easy half of the problem: connect the thing once you already know its name. That was always a little too neat for real enterprise systems, where nobody knows the name and half the tools live somewhere inconvenient. ARD looks less glamorous, which is usually how the actually useful standards start. A spec for finding the tool before you worship it is a sensible bit of grown-up plumbing, and the industry could use more of that and less victory-lap architecture.
Read more about this at: The New Stack
