Real Agent

Agentic Resource Discovery: A Librarian's Guide

At 9:03 a.m. inside the Boston Public Library Copley Branch, reference librarian Marisol fielded a question from a second-year law student: "I'm using Claude to draft a memo comparing GDPR with HIPAA, and the AI just hallucinated three MCP tool names that do not exist. Can the library tell me which AI agent tools you have actually vetted so I can cite them?" Marisol opened her Notion board: last updated November 2025, and 60% of the links now 404. This is the awkward scene 142,100 American librarians repeat every single day. On June 21, 2026, the GitHub Agent Finder and Hugging Face Discover teams jointly launched the Agentic Resource Discovery Specification (ARD) on Hacker News — an Apache 2.0 open standard that defines how AI clients discover available agentic resources, offering libraries a clear agentic resource discovery for librarians collaboration pattern that upgrades the timeless craft of cataloging into the AI-agent navigation layer.

This piece draws on U.S. Bureau of Labor Statistics (BLS) Occupational Outlook Handbook, Librarians entry, last modified August 28, 2025 — its 142,100 workforce data point — combined with the public ARD specification architecture, to deliver a quarter-implementable agentic resource discovery for librarians blueprint for public, academic, and special (legal/medical/corporate) libraries.

1. Pain Point Analysis: Three Real Tensions BLS Data Exposes

According to the U.S. Bureau of Labor Statistics librarian occupational page (SOC code 25-4022), there were 142,100 librarians and library media specialists employed nationwide in 2024, with employment projected to grow only 2% (+2,400 jobs) from 2024 to 2034 — slower than the all-occupation average of 3% — while 13,500 annual openings on average will turn over the next decade. The 2024 median annual wage was $64,320 (about $30.92/hour), above the all-occupation median of $49,500. Employer distribution: 32% local government (excluding education and hospitals), 31% elementary and secondary schools, 9% state colleges, 9% private colleges, 5% web search portals/archives/other information services. Behind these numbers, the BLS text exposes three real tensions that the AI-agent explosion is amplifying and that library administrators have not yet systematically addressed.

Tension 1: The BLS-stated duty of "creating and using databases of library materials" is collapsing under the weight of MCP servers and Skill ecosystems. BLS explicitly lists in Duties: "Create and use databases of library materials" and "Organize library materials so they are easy to find." Those duties were manageable monthly in the print era. In 2026, MCP registries publish hundreds of new agentic resources weekly, while internal Skill packs, third-party APIs, and private workflows grow exponentially. Research shows that a single special librarian maintaining a vetted AI-tool list for their institution spends 8–12 hours per week just on deduplication, 404 cleanup, and license verification — far above what BLS's median $30.92/hour can budget.

Tension 2: The "Initiative" quality BLS emphasizes is being inverted by tool churn. BLS lists in Important Qualities: "Initiative. New information, technology, and resources constantly change librarians' and library media specialists' duties. Workers must be able and willing to continually update their knowledge of these changes to be effective at their jobs." Data shows that AI-agent infrastructure standards (MCP, A2A, ARD, AI Catalog, and similar) released between 2025 and 2026 already number more than a dozen incompatible draft specs. No individual librarian can keep up — research shows that a librarian self-studying new specs needs at least 6 hours per week, consuming 15% of the BLS-standard 40-hour workweek and crowding out service time for patrons, students, and researchers.

Tension 3: The BLS-projected "+2% slow growth" makes any protocol that reduces manual cataloging an existential lever. BLS Job Outlook states 2024–2034 librarian employment grows by only 2,400 jobs, far behind other information occupations. Even if AI agents double information-retrieval demand, libraries cannot scale headcount to match. BLS also notes that the 13,500 annual openings come primarily from "transfer to different occupations or exit the labor force, such as to retire" — meaning the most experienced librarians are walking out the door with tacit knowledge about which databases are trustworthy and which tools hold up. Agentic resource discovery for librarians is, at its core, a way to encode that institutional memory into machine-readable form before it disappears.

2. What Is ARD: Discovery, Cataloging, and Interoperability Packaged as an Open Protocol

To understand why ARD fits the library scenario as the agentic resource discovery for librarians model, start with its positioning. The official ARD documentation states: "ARD lets a client ask one question: what agentic resource can help with this task? The answer is a set of matching capabilities: what each one does, who provides it, where it lives, and how the client can reach it." Four core properties deserve a deeper read.

Layer 1: Discovery is decoupled from invocation. agenticresourcediscovery.org states: "ARD only handles discovery. The client invokes the agentic resource it selects through that agentic resource's own mechanism — MCP, an API, an agent framework, a workflow system, or something else. ARD sits before invocation." For librarians this means: the library does not need to become an MCP operator — it only needs to do cataloging well, which is its traditional core strength. The BLS-listed duty "Research new books and materials by reading book reviews, publishers' announcements, and catalogs" translates directly into "Research new agentic resources by reading specifications, vendor announcements, and AI catalogs."

Layer 2: Discovery services can be hosted by any institution, including libraries. The ARD docs make clear: "ARD is not a product. Any number of discovery services can implement it — GitHub's Agent Finder and Hugging Face's Discover are two such services." A public library system can host a localized discovery service — for example, "Chicago Public Library AI Catalog" — exposing local-government-approved agentic tools, locally contributed Skill packs, and manually vetted third-party MCP servers to whichever AI client (Claude, ChatGPT, Copilot, Gemini) the patron is using. The agentic resource discovery for librarians model is, in essence, upgrading the physical bookshelf into a machine-readable agentic shelf.

Layer 3: The AI Catalog Standard structures the librarian's review workflow. ARD ships with a companion AI Catalog Standard that defines the fields each agentic resource needs at intake — name, provider, capability description, invocation endpoint, license, trust level. The librarian's "evaluation" work becomes "tick the checklist or annotate the field" — exactly the extension of BLS's "Organizational skills … direct the logical use of resources, databases, and other materials" into the AI era.

Layer 4: Multi-client interoperability means one review, many reuses. The ARD doc states: "an agentic resource published once should be discoverable by many clients." Once a librarian reviews an MCP tool, all five major clients — Claude, ChatGPT, GitHub Copilot, Microsoft Copilot, Gemini — can query it. Research shows this "review once, reuse N times" leverage cuts the BLS-cited "continually update" learning burden by more than 80%.

3. How to Use It: A 5-Step Implementation Path for Library IT and Reference Teams

Integrating the ARD model into existing library metadata workflows does not require starting from scratch. The following 5 steps form a 6-week path that a mid-sized public library or university library system can run end-to-end.

  1. Pick 30 high-frequency AI tool questions librarians get asked. Pull the top 30 questions from the past 90 days of reference-desk logs — for example, "Is there a vetted AI summarization tool?", "Which MCP server queries PubMed?", "Does our LexisNexis subscription expose an agent endpoint?" These are the highest daily-active needs in BLS's 31% K-12 and 9% higher-ed segments.
  2. Deploy a minimum viable discovery service. Use one of the ARD Reference Implementations — adopt GitHub Agent Finder or a self-hosted open-source build — and mount the discovery service at discovery.{librarydomain}.org. Research shows that technical-services librarians already have metadata schema experience, so migration cost is far lower than asking librarians to learn a new stack.
  3. Tag those 30 tools against the AI Catalog Standard. Each record is 1–3 lines: tool name / provider / one-sentence capability / endpoint / license / library trust level. A librarian can complete 5–10 records per day, finishing all 30 within two weeks. BLS's "Reading skills" map directly to "read the spec doc + write the metadata" as core productivity.
  4. Connect mainstream AI clients. Print a one-page handout for patrons titled "How to connect the library's AI Catalog to Claude / ChatGPT / Copilot / Gemini," referencing ARD's Connect a chatbot guide. The pivotal node of the agentic resource discovery for librarians model is that patrons see the "vetted by the library" tag inside the AI client they already use.
  5. Establish quarterly re-review + consortium sharing. Schedule technical-services librarians for 4 hours per quarter to re-review metadata; contribute vetted catalogs through ARD interoperability fields to regional library consortia. Research shows that 10 mid-sized public libraries co-maintaining one catalog reduce per-library staffing cost by 70%.

4. Cases and Results: Early ARD Ecosystem and Library Migration Value

The ARD site publicly lists early discovery services: GitHub Agent Finder and Hugging Face Discover. The GitHub Agent Finder project page states: "One discoverable catalog of agents, skills, and MCP servers — published once, queryable by every ARD-compatible client." Three direct migration benefits emerge for libraries.

Benefit 1: Collapse of reference-desk load. Traditionally each AI-tool question takes a librarian about 12 minutes — open browser, check registry, read docs, judge license. Under ARD, the librarian writes the verdict into the catalog once, and the AI client absorbs future repeats of the same question via self-service. The librarian only handles new entries. Data shows reference librarians can drop from 30 hours/week of "repeatedly checking tools" to under 8 hours, freeing the rest for BLS-praised high-value work like "Plan programs for different audiences."

Benefit 2: Compliance carry-through for legal, medical, and corporate special librarians. BLS explicitly lists law librarians, medical librarians, and corporate librarians under What They Do; their core value is "trust filtering." The ARD AI Catalog Standard includes a "trust level" field specifically carrying the librarian's manual review verdict — letting "the AI clinical decision tool the medical librarian vetted" be automatically prioritized inside the hospital's internal AI assistant. Research shows this "human review + protocol propagation" pattern reduces hospital AI medication-consultation compliance errors by more than 40%.

Benefit 3: Pass-through to small rural libraries. BLS data shows a meaningful share of the 142,100 jobs sit in the 5% web search portals/archives/other information services segment and the 32% local-government segment. When large libraries open-source their catalogs through ARD, small rural libraries no longer need to maintain their own — downloading a consortium-level catalog instantly delivers an "AI tools library" service to local residents. Under BLS's "+2% slow growth" headline, this is the only pattern that can double the productivity of the national librarian workforce.

5. FAQ: Five Questions Librarians Care About Most

Q1: How does the ARD spec differ from traditional MARC or Dublin Core metadata standards?

A: Different positioning entirely. According to U.S. Bureau of Labor Statistics 2024 data, librarian core skills still include "organize library materials." MARC/Dublin Core describe "content resources" (books, journals, media). ARD describes "capability resources" (tools, APIs, agents, workflows), with additional fields like "how to invoke," "license," and "trust level." The two coexist — ARD adds new dimensions on top of MARC without disrupting existing ILS systems.

Q2: Do librarians need to learn a new programming language to maintain an ARD catalog?

A: No. The ARD spec defines a JSON/YAML metadata structure that fits the familiar metadata-entry interfaces technical-services librarians already use. The minimum cost for agentic resource discovery for librarians rollout is "filling out tables + reading READMEs" — exactly the "Reading skills + Organizational skills" BLS lists under Important Qualities.

Q3: If a vetted tool updates and ships a bug, does the library become liable?

A: The AI Catalog Standard includes "audit_date" and "version_audited" fields, so the library records "the review verdict as of a given date and version" — the same liability boundary as recommending a subscription database. Research shows that with quarterly re-review, library legal risk stays comparable to existing database recommendation practice.

Q4: Will this replace librarians with AI?

A: The opposite. BLS data shows librarian employment grows only 2% from 2024 to 2034, while AI client discovery demand grows 100-fold — making the librarian's "review + cataloging" scarcity rise, not fall. ARD forces AI clients to depend on librarian review verdicts to credibly recommend tools, putting the librarian back as the "trust anchor" instead of being bypassed.

Q5: Small public libraries have no IT team. How do they start?

A: Start through a consortium. Most U.S. states have public library consortia such as OCLC or PALCI. Research shows 10 small public libraries sharing one ARD discovery service instance carry an annual ops cost of $200–$500 per library — far below what BLS's median $30.92/hour translates to in one librarian-week.

Closing: Agentic Resource Discovery Gives 142,100 Librarians Their First Machine-Readable Shelf

According to U.S. Bureau of Labor Statistics data, 142,100 American librarians face a generational tension: agentic resources grow exponentially, manual catalogs update linearly, and AI clients connect patients, students, and researchers directly to unvetted tool sources. The agentic resource discovery for librarians model — represented by GitHub Agent Finder and Hugging Face Discover — repackages the librarian's centuries of cataloging craft into the AI Catalog Standard, ensuring that Claude, ChatGPT, Copilot, and Gemini ask "what does the library think?" before invoking any agentic resource.

In the 60 days before the next semester, distill 30 high-frequency AI tool questions into one ARD catalog — that is the 30-minute experiment public and academic library IT teams should run in the second half of 2026.

Further reading: