The discovery layer breaks: agents don't browse your course catalog

For two decades, the discovery problem in education and credentialing was solved the same way: a catalog, a search box, and marketing copy. A human browsed. They read course descriptions, compared programs, absorbed brand signals, and made a choice. Every layer of the stack — SEO, landing pages, enrollment funnels, glossy hero images — was optimized for one act: a human deciding to trust an institution enough to enroll. That model is breaking. Not because catalogs got worse, but because the entity doing the discovering is changing. Agents don’t browse.

The catalog was never the product. The interface was

A course catalog is an interface. It is a structured surface designed to convert a human’s vague intent (“I want to learn machine learning”) into a specific enrollment decision. The interface works because humans are pattern-matching on trust signals: institutional brand, accreditation logos, testimonials, curriculum structure, price. The catalog is a trust-conversion machine. Every pixel is engineered for human attention.

Agents operate on a different logic. An agent doesn’t scroll. It doesn’t convert on emotional copy. It queries, resolves, and composes. When an agent encounters a credential or a learning pathway, it wants to answer a small set of hard questions: What does this credential certify? Who issued it? Is it still valid? What evidence backs the claim? Can it compose with other credentials into a coherent pathway? Can I verify it in real time?

A catalog optimized for human browsing answers none of these questions in a machine-resolvable form. The course page has a marketing description of the learning outcomes. It does not have a signed, verifiable claim linking a learner to demonstrated competencies. The institution’s accreditation is listed as a badge image. It is not resolvable to a live registry. The credential itself, once issued, lives in a PDF or a proprietary wallet that renders as a visual display — designed for a human to screenshot, not for an agent to parse.

The catalog built for humans is the wrong interface for agents. And as agents mediate more discovery — recommending pathways, assembling learning plans, evaluating candidate qualifications — the catalog becomes invisible. Not deprecated. Invisible. Agents route around surfaces they can’t resolve.

The gap nobody has mapped

Here is the gap, plainly: agent-mediated discovery is being theorized almost entirely on the consumer side. Every, Stratechery, Latent Space — the serious strategy outlets are writing about agents discovering products, content, and services. The frame is demand-side: how will agents find your product, how will they evaluate it, what interface do they need. That conversation is live and getting sharper.

On the education side, EdTech analysts are still writing about catalog discovery for humans. SEO for course pages. Enrollment funnel optimization. The credential wallet as a display surface. The frame is supply-side but human-bound: how do institutions present credentials to learners and employers who will read them.

Nobody is mapping agent-mediated discovery onto credentialing supply. The question is not asked: when an agent queries a credential, what does it need to resolve, and who is building the interface that lets it resolve it? The demand-side analysts don’t build credentials. The supply-side analysts don’t build for agents. The gap is real and it is structural.

We can map it because we are building the supply side. Mneurix Lattice is an evidence-first, agent-readable credentialing system. The design decisions we have already made — signed verifiable credentials, issuer DIDs, live revocation endpoints, evidence pointers, composability — are not features for a future market. They are the interface agents need today, and almost no one is providing it.

Credentials-as-evidence, not credentials-as-marketing

There are two models of what a credential is, and they are diverging fast.

The marketing credential is a brand signal optimized for human trust. Its job is to signal quality to a human viewer: a logo, an institution name, a recognizable badge, a certificate frame. The marketing credential gains its power from institutional reputation and the viewer’s pattern recognition. It is a trust shortcut. It works because humans are good at approximating trust from brand signals and bad at verifying claims directly.

The evidence credential is a machine-resolvable bundle optimized for agent verification. Its components are not visual. They are structural: a signed verifiable credential, an issuer DID, a live revocation endpoint, and evidence pointers — links to the work, assessments, or demonstrations that back the credential’s claim. An agent can parse this bundle, verify the signature, check revocation status, inspect the evidence, and compose the credential with others into a coherent picture of what a learner actually knows and can do.

As discovery moves to agents, the marketing credential loses leverage. A brand badge means nothing to an agent that can’t resolve it. The evidence credential gains leverage because it is the only form an agent can actually use. The catalog that serves agents serves evidence, not copy.

This is not a theoretical distinction. We are already building it. A Mneurix Lattice credential is not a PDF with a logo. It is a signed, resolvable, evidence-linked object. When an agent queries it, it gets answers. When an agent queries a traditional credential, it gets a wall.

Discovery and custody are the same stack

If the agent-readable credential interface is the aggregation point, then the question is: what makes a credential credibly resolvable to an agent in the first place?

The answer is custody. We wrote about this in The custodian problem: a credential is only as trustworthy as the custody of the issuer key. If an issuer’s signing key is loosely held — shared across systems, stored without hardware-backed protection, recoverable through a support ticket — then the signed credential is a soft claim. An agent verifying that credential has to discount it. The signature is technically valid but operationally weak. The issuer’s key custody is the load-bearing wall.

This is why discovery and custody are not separate concerns. They are the same stack. An agent resolving a credential needs two things: a resolvable interface (signed VC, DID, revocation endpoint, evidence) and credible custody behind the signature. A credential with a beautiful agent-readable interface but thin key custody is a well-structured lie. A credential with strong custody but no agent-readable interface is a locked vault. You need both.

The implication for aggregation is direct. Whoever owns the agent-readable credential interface — the layer that resolves credentials fastest, most completely, and with the strongest custody guarantees — becomes the aggregation point. Agents will route through the layer that gives them the best resolution. This is the same dynamic Ben Thompson described in aggregation theory: the aggregator owns the user relationship because it owns the best interface to the underlying supply. Here, the “user” is the agent, and the “supply” is verifiable credentials. The aggregator is the layer that makes credentials resolvable.

We covered the aggregation logic in Aggregation theory for credentials. The addition is that custody is what makes the aggregation work. Without credible issuer key custody, the resolution layer is passing through credentials that agents can’t fully trust. The aggregation point is not just the fastest resolver — it is the resolver whose credentials are backed by sealed-key custody at the issuer level. That is the stack we are building: agent-readable interface on top of evidence-first credentials on top of sealed-key custody.

The land-grab window

This is time-sensitive. The agent-mediated discovery layer is not a future market — it is forming now, and the land-grab window is measured in months, not years.

The incumbents — LMS catalogs, credential wallets, EdTech marketplaces — are structurally misaligned for what’s coming. Their entire surface is optimized for human attention: conversion funnels, copywriting, visual hierarchy, call-to-action buttons, PDF previews. That’s what a course catalog is: a trust-conversion interface layered over a credential supply that was never built to be resolved by anything except a human eye. When an agent hits that surface, it finds no machine-resolvable endpoints, no signed issuer DIDs, no revocation status, no evidence pointers. It finds a marketing page. The agent can’t compose, can’t verify, can’t resolve. It fails silently and routes elsewhere.

The incumbents won’t re-architect cheaply. Their catalogs are deeply coupled to enrollment flows, revenue models, and institutional politics. Decoupling the credential supply from the human presentation layer is not a redesign — it’s a re-platforming, and the organizations most invested in the current catalog form are the least likely to undertake it. They treat agents as a future problem. That’s the window. It stays open exactly as long as they keep treating agent-mediated discovery as speculative, as something that lands after the next LMS release cycle, after the next standards committee, after the next budget round. Meanwhile, the layer that resolves credentials for agents is being built by people who don’t own catalogs — and that asymmetry is the opportunity.

The defensible position is being the evidence-first, agent-readable layer before the incumbents notice — and that position compounds. Here’s the mechanism: agents route to whichever layer resolves credentials best. Resolution quality is the entire product from the agent’s perspective. An agent that queries for “verified competence in distributed systems” doesn’t care about your brand, your enrollment funnel, or your institutional prestige. It cares whether the credential it receives is signed by a custody-rooted issuer, resolvable to a live revocation endpoint, and linked to evidence it can follow. The layer that provides that resolution reliably gets routed to again. That routing is sticky — it’s aggregation dynamics, the same pattern I mapped in the custodian problem. Once an agent’s resolution path is habituated to a particular aggregator, that aggregator owns the interface. Competing layers have to not only match the resolution quality but dislodge the routing, and dislodging routing means convincing agents (and the pipelines that build on agents) to switch. First-mover on the agent interface is the durable position. Latecomers face a brutal choice: un-build their human-optimized surfaces to compete on resolvability, which they won’t do fast, or build a parallel agent-readable surface that nobody routes to, which they won’t do well. The economics favor the layer that was agent-readable from the start.

Now the honest caveat, because overclaiming here is worse than silence. Agent-mediated credential discovery is early. Genuinely early. There is no agreed agent↔credential protocol — no RFC, no working group consensus, no de facto standard that the way agents query credentials is how they’ll query them in two years. The agent space itself is volatile: frameworks shift, evaluation benchmarks are contested, reliability at interpreting credential evidence is uneven and sometimes poor. An agent that correctly resolves a signed VC today may mishandle the same VC next quarter if its tooling changes. The bet is directional, not precise. The claim is that the supply side should be agent-readable now because that’s unambiguously where discovery is going — the catalog is not coming back, human-mediated browsing is not the future of credential discovery at scale — but the exact interface, the exact protocol, the exact way an agent expresses “give me the evidence for this claim” is not settled. It may settle through W3C, it may settle through an agent framework becoming dominant, it may settle through a de facto aggregator’s API. We don’t know which. What we know is the direction. Building evidence-first credentials — signed, resolvable, revocable, evidence-linked — is correct regardless of which interface wins, because every plausible interface requires exactly those properties. The properties are invariant; the wire format is not. That’s the bet: build the invariant layer, stay portable on the variable one.

The discovery layer is breaking along the seam between human-optimized and agent-resolvable. It is breaking now, and the break is permanent. The catalog does not come back. The credential that survives the break is the one an agent can verify end-to-end — signed by a custody-rooted issuer, resolvable to live revocation, linked to evidence the agent can follow and a human can inspect. Everything else is a marketing page that no one — no agent, and increasingly no human — resolves to. Build that credential. Make it resolvable. Don’t wait for the catalog to come back. It won’t.