Executive Summary
Agent rollouts are hitting a wall that has nothing to do with model quality. Every tool an agent can reach was wired in by hand, and that list goes stale the moment infrastructure changes. The missing layer is discovery.
Agentic Resource Discovery, or ARD, is an open spec built to close that gap. It lets an agent search registries for the resource a task needs instead of being handed a map. The Model Context Protocol settled how a model calls a function. It never settled how the agent learns the function exists. That omission gets expensive at volume, where hundreds of agents reach for thousands of resources across clouds, vendor catalogs, and team indexes. Registries double as a control point, so publishing what exists and deciding what an agent may see collapse into one job. Watch whether ARD stays open. A discovery layer only works when any vendor can register and any agent can search.
An agent shows up on the job already knowing too little about what you actually run. It carries a list of tools someone typed into a config file, a few endpoints hardcoded in the prompt, and no way to discover anything else. Agentic Resource Discovery intends to change that by standing up a DNS for the agent era.
The idea is simple. Instead of telling an agent where every resource lives in advance, you let it query a registry to find what it needs. If DNS turns a name into an address, ARD turns a need into a resource. The agent searches across registries rather than waiting for you to hand it a map.
Pre-configuring agents breaks down at scale
Most agent setups today are wired by hand. Someone curates a list of tools, writes the connection details into the system prompt, and the agent only ever sees what that list contains. This works fine in a demo. It collapses the moment you have real infrastructure.
The breaking point is volume. An organization with hundreds or thousands of resources scattered across clouds, platforms, and internal services cannot maintain a hand-written inventory that stays current. Each new tool adds wiring. Each rotation of a credential breaks something. Each resource nobody thought to mention is simply invisible to the agent.
That last part is the real cost. An agent cannot use a tool it does not know exists. The moment you stop maintaining the list, you have built a ceiling on what your agents can actually do.
Drift is the quiet version of this. Teams rarely decide to abandon the inventory. They just stop updating it. New services launch, old endpoints get retired, and the agent keeps working from a stale picture. Over months the map drifts far enough that an agent pointed at a dead resource returns an error, which is worse than never knowing the thing was there.
What ARD adds
ARD is an open specification rather than a product. It defines how an agent searches across registries for the resources it needs. In this model the agent brings a question about its task, and the registry answers with the resource that fits. Discovery becomes a service instead of a setup chore.
The contrast with current practice is stark. Hardcoding a tool into a prompt is brittle. Searching for it is adaptive. An agent that can ask what is available for a task can handle a new resource the moment it gets registered, with no redeploy and no changed prompt.
The spec is also built for federation. Resources do not live in one place in a large enterprise. They sit in public vendor directories, internal catalogs, and team-specific indexes. ARD anticipates that spread by letting an agent query several registries rather than one central list. The agent stays free to find what it needs wherever it is published.
Why this intersects with the wider agent movement is that a major cloud player called attention to it in a hosted weekly roundup. When a big vendor surfaces a protocol, it usually means the company sees a gap that needs a shared answer. The gap here is that nobody has defined how an agent learns what it can call.
Registries also double as a control point. Publishing to one means setting out what exists, what it does, and how to reach it. That same surface is where you decide what an agent may see. Discovery and governance end up in the same place, which is a real advantage for teams nervous about letting agents loose.
The missing layer in agent tooling
Agent tooling has spent a couple of years wiring the interface between a person and a tool. Function calling and the Model Context Protocol let a model call a function cleanly. What neither solved is how the agent learns the function exists in the first place. The interfaces got fixed and the discovery problem quietly went untouched.
ARD is an early and honest attempt to close that gap. It only matters if it stays open. The value of a discovery layer comes from a shared set of registries where any vendor can register a resource and any agent can search it. That is precisely what made DNS work, and it is the only version of this that scales past one vendor’s tooling.
Hard questions remain. There is the matter of who runs the registries, how they authenticate, and how an agent trusts what it finds. There is the need for a common descriptor so resources can be compared and consumed without bespoke integration per vendor. None of these are trivial. But for the first time someone has named the problem and published a spec that tries to solve it.
The takeaway is that the next constraint on your agent rollout is not model quality or tool quality. It is discovery. The organization that can let an agent find a resource on demand will support far more automation than the one that must hand it a list.
Related reading. Treat Your AI Agent Rules Files Like the Code They Replace. Cortex Builds an Internal Developer Portal That Answers Your Own Questions. The Cost of AI Is Finally Falling. The Cost of Using It Is Not.. AI Infrastructure Runs on Four Layers. Most Break Below the Model..
Get the next one before it is old news
Independent analysis of cloud-native infrastructure, Kubernetes and data centre economics. No vendor spin.
[…] reading. We wrote about giving agents their own resource discovery layer, about the discovery gap MCP left open, and about why agent rule files are […]