NEW SPEC LETS AI AGENTS DISCOVER TOOLS WITHOUT MANUAL SET UP
The number of external capabilities for AI agents is exploding, but finding and connecting them is still done by hand. ARD aims to fix that.
by editor6 min readcomments soon
The number of things an AI agent can reach for is growing fast: APIs, MCP servers, workflows, Skills, and other agents. All of these get called agentic resources, and they represent the difference between a chatbot that only knows what it was trained on and one that can book a flight, query a database, or control a smart home.
But there is a problem. Finding and using those resources is mostly manual. A developer or an IT admin has to find each tool, judge whether it is trustworthy, wire it into the client, and maintain the connection over time. That works for a handful of well-known tools, but it breaks down as soon as dozens or hundreds of entities publish their own agentic resources.
The bottleneck is no longer invocation. It is discovery.
WHAT ARD DOES
The Agentic Resource Discovery Specification (ARD) was written to solve that specific bottleneck. Instead of having a user or developer hunt for resources, ARD lets an AI client simply ask what agentic resource can help with a given task. The specification only handles the discovery part: once the client finds the right resource, it invokes it through that resource's own mechanism, whether that is a REST API, an MCP protocol, or something else.
ARD is not a product. Any discovery service can implement the specification. GitHub's Agent Finder and Hugging Face's Discover are two examples of services that could implement the spec. The design is deliberately open: a resource published once should be discoverable by many clients, letting agents reach beyond the small set of tools they have been pre-configured to know.
WHY DISCOVERY IS THE HARD PART
Initially, the set of available tools was small. A team might wire up a weather API, a calendar, and a search engine, and that was good enough. Now the landscape looks different. More entities are publishing agentic resources. Open-source projects are shipping MCP servers by the week. The ecosystem is fragmenting into hundreds of specialised capabilities, and no developer wants to hand-craft a catalogue entry for each one.
Manual discovery does not scale. It also creates a security problem: when a user has to manually vet every resource, they either vet none of them or they fall back on a curated list that misses useful alternatives. ARD moves the vetting burden to the discovery service, which can maintain reputation, freshness, and access controls on behalf of many clients.
The specification is designed to be lightweight. It defines a query interface: given a natural language description of a task, what resource can help? The client gets back a URI, metadata, and an authentication hint. Then it proceeds on its own. No central registry, no mandated format for the resources themselves, no new protocol to implement.
NEW DOORS OPEN
For a developer building an agentic application, ARD removes the need to hardcode a list of tools. The agent can ask the discovery service at runtime and get back a list of available email APIs with their authentication requirements. That same agent can work in different contexts with different discovery services, or fall back to a local cache when offline.
For an end user, the experience should become invisible. Instead of installing plugins or granting API access to every service separately, the user's AI client consults a discovery service, negotiates authentication once, and proceeds. The user only sees that the agent suddenly knows how to do things it could not do before, without manual setup.
The specification also opens the door to marketplaces of agentic resources. A company that publishes a useful API can make it discoverable through ARD-compatible services, and any agent using those services can find it. That lowers the distribution cost for tool builders and gives agents a richer set of capabilities than any single vendor could provide.
AUTONOMOUS ASSISTANCE IS HERE
ARD is one piece of a larger shift. AI agents are moving from executing single commands to orchestrating multi-step workflows that involve several tools and data sources. Each of those steps requires a discovery step: what tool can parse this PDF, which database holds the customer's order history, and which API can schedule the delivery. Without a standardised way to ask those questions, every agent builder has to reinvent the discovery layer.
An open specification like ARD creates a common language. A client built by one team can discover resources published by another team, across companies, across ecosystems. That interoperability is what makes the agent metaphor more than a demo. It means a user can hire an agent to plan a vacation, and the agent will discover a flight booking tool, a hotel search API, a weather service, and a local recommendations engine, all without the user installing anything.
The spec does not solve every problem. Authentication, authorisation, rate limiting, and cost tracking are left to the resource's own mechanisms. ARD is narrowly scoped: it answers the question and hands the rest to the client. That narrowness is intentional. It keeps the specification simple enough that many providers can adopt it quickly.
WHO IS IT FOR THOUGH
GitHub and Hugging Face are early examples. GitHub's Agent Finder indexes open-source agentic resources from repositories, and Hugging Face's Discover does similar work on the model and dataset side. Both have the user base to make discovery meaningful. Other platforms could implement ARD-compatible discovery without much overhead, turning their existing registries into agentic resource directories.
The real test will be whether enterprise SaaS providers publish their APIs with ARD metadata. If they make their endpoints discoverable, agents built for business automation can find them without per-supplier integration work. That is where the value compounds.
NOAM SHAZEER JOINTS OPENAI
In a separate development, Noam Shazeer has joined OpenAI. While not directly tied to the ARD specification, the hire signals that the talent race for building sophisticated agent systems is accelerating alongside the infrastructure race for making those agents easy to operate.
ARD IS STILL EARLY
ARD is still early. The specification needs adoption to matter, and adoption depends on both discovery services and resource publishers supporting it. The initial implementations from GitHub and Hugging Face give it a critical mass. If the pattern holds, expect more discovery services to pop up, along with registries that curate resources by domain, quality, or cost.
The manual discovery era for agentic resources is ending. ARD is the attempt to write a clean transition to the next stage, where agents find what they need without a developer holding their hand. The pieces are there. Now they need to connect.
what did you make of it?
more from ai
ai
TSMC ADDS $100 BILLION TO ARIZONA CHIP BET, TOTAL HITS $265 BILLION
The additional investment will build at least four more 2nm fabs and advanced packaging, bringing the company's total US commitment to $265 billion.
ai
META WILL ALERT PARENTS IF TEENS DISCUSS SUICIDE WITH META AI
The opt-in feature flags self-harm references in chatbot conversations, with human review before any notification is sent.
ai
ROBLOX'S "BUILD" LETS ANYONE MAKE A GAME FROM THEIR PHONE WITH AI
The new toolset, launching July 28, turns text prompts into playable experiences and puts game creation on iPhone and iPad.
ai
ZOOX REALLS ENTURE ROBOTAXI FLEET OVER SMOKE DETECTION FAILURE
A robotaxi drove into an active fire scene obscured by smoke. NHTSA called emergency scenes not edge cases.