The Agent Discovery Gap

The agentic web has increasingly sophisticated ways for software to find other software. MCP lets agents connect to tools. A2A lets agents talk to agents. ANS lets agents verify identities. What none of them can do is help an agent find a verified human expert.

This is the most important unsolved problem in the agentic web — and the one with the highest stakes for getting right.

The Consequences Are Already Visible

The gap is not theoretical. A $290,000 Deloitte report contained approximately 20 fabricated references, including a nonexistent book attributed to a real Sydney University professor. A Stanford professor — recognized for his expertise on AI and misinformation, of all things — had his sworn expert testimony excluded from federal court because his AI-generated citations were fake. Therapy chatbots on Meta fabricated detailed credentials, with one bot providing the real license number of a Maryland counselor who was completely unaware.

These are not edge cases. They are the predictable result of a system where AI determines "expertise" through proxy signals — who sounds authoritative — rather than through verified credentials.

How AI Actually Finds "Experts" Today

None of the major AI platforms operate a formal expert identification system. They find authoritative content, not verified people. Perplexity runs a proprietary index of 200+ billion URLs. ChatGPT delegates retrieval to Bing. Claude uses Brave Search, with 86.7% overlap between its cited sources and Brave's top results.

All three determine expertise through the same proxy signals: institutional affiliation mentioned in text, publication frequency, backlink profiles, engagement metrics. The Princeton GEO study quantified what works: adding verifiable statistics boosts AI visibility by 40%, citing credible sources adds 40%, authoritative tone contributes 25%. These are content-formatting signals, not expertise verification. A well-optimized page by a non-expert can outrank a genuine specialist who publishes without these structural elements.

The Protocol Ecosystem Has a Human-Shaped Hole

MCP, donated by Anthropic to the Linux Foundation, has achieved 97 million monthly SDK downloads and over 10,000 active public servers. A2A, contributed by Google, has 150+ participating organizations. ANS creates a DNS-inspired registry with PKI certificates. Every one of these protocols defines discovery for software entities, not for human experts.

The core problem: A2A Agent Cards describe what an AI agent can do. MCP Server Cards describe what tools a server exposes. Neither has a way to represent that a board-certified specialist with 20 years of experience is available for consultation. The protocols are rails for moving information between machines. What is missing is the context — who the expert is, what they know, and whether it is worth querying them.

This distinction is explored further in the deep dive on MCP and A2A as rails, not context, and the conference booth analogy describes what expert-agent interaction should actually look like.

What a Solution Requires

Five layers would need to exist. First, a standard for Expert Agent Cards — extending A2A's Agent Card with verified credentials, expertise domains, and communication endpoints. Second, automated credential verification that aggregates signals from licensing boards, ORCID, and professional registries. Third, query routing with compensation, enabled by payment protocols like x402. Fourth, trust and reputation scoring that combines credentials, publications, and peer endorsement. Fifth, compliance automation — the agent equivalent of the confidentiality checks and conflict screening that firms like GLG maintain for their human-mediated expert consultations.

The components exist in fragments. The NPI Registry lets anyone verify a healthcare provider's license in real time. ORCID provides machine-readable profiles for 18 million researchers. W3C Verifiable Credentials v2.0 enables cryptographically tamper-proof credential exchange. But these pieces have never been assembled into a coherent discovery stack.

The Stakes Are High and Getting Higher

The economic case is substantial. Gartner projects 25% less traditional search by 2026. McKinsey estimates $3-5 trillion influenced by agents by 2030. Whoever builds the integration layer between AI agents and human expertise captures a position analogous to what Google captured between intent and information — except that in this market, errors are measured in misdiagnoses rather than irrelevant search results. The broader context around the AI content crisis and how expertise sharing is changing explains why this gap matters now.