The Layer Everyone Is Reinventing

Library science, world model research, and agentic engineering each arrived at the same missing infrastructure in 2025–2026, from opposite directions, without reading each other. eBay proved it worked: +89.8% retrieval recall by routing through an explicit taxonomy instead of embedding space. LeCun's H-JEPA discovered the same structure from the physics of prediction. Agentic engineers call it a skill library. Three names, one diagnosis: a system that must act over a domain has to organize what that domain contains — explicitly, outside the weights, in a structure that can be queried without retraining.

The pattern that keeps appearing in this reading queue is disciplines discovering the same thing. The semantic infrastructure crisis — six authors, zero citations of each other, one diagnosis — ran four years ago. The agent memory problem — six perspectives, three tensions, no overlap — ran eight months ago. Now a third version of the same thing is playing out, and it involves a more surprising cast: a semantic engineer with twenty-five years in library science, the Yann LeCun research program on Joint Embedding Predictive Architectures, and the cohort of engineers in June 2026 who have been running agent loops long enough to have opinions about what makes them fail.

Jessica Talisman's essay "Taxonomies, All the Way Down" is the piece in this batch that earned the longest read. The argument is dense and the citations are real — Rosch, Berlin, Collins and Quillian, Collins and Loftus; the cognitive science of categorization from the 1970s onward. But the empirical anchor that makes it land is more recent. eBay replaced embedding-similarity search for product classification with LLM-guided taxonomy traversal and measured a 89.8% improvement in F1, an 86.1% improvement in precision, and a 72% improvement in retrieval recall. The traversal visited only 1.7% of taxonomy nodes per query. The system that knew what it was looking for — had an explicit structure to route through — was not faster than the system that did not; it was qualitatively different, in the same way that a reference librarian who knows the catalog is different from one who skims shelves.

The eBay result is not a trick of architecture or a cherry-picked benchmark. It is a precise demonstration of the gap this site has been circling from the data side: explicit structure beats approximate retrieval over a domain. For product classification, that structure is a taxonomy. For governed data, it is a semantic layer with typed context. The word changes; the mechanism is the same.

1. What Talisman is actually claiming

The essay's headline thesis is that every significant layer of AI infrastructure — retrieval, grounding, risk governance, world models, agentic routing — converges independently on hierarchical taxonomic organization. This is a strong claim, and Talisman earns it in three distinct moves.

The first is representational: LLMs already encode taxonomy-shaped structure in their weights. The 2024 ICML result she cites found that categories organize into polytopes in representation space, with hierarchy expressed as near-orthogonal directions at different abstraction levels. The structure is there. It is also implicit, unaddressable, and non-revisable without retraining — which is why an explicit taxonomy outside the weights is not redundant but complementary. The external structure solves the three problems the internal structure cannot: you can query it, update it, and attribute answers to it.

The second move is architectural: Yann LeCun's H-JEPA (Hierarchical Joint Embedding Predictive Architecture) explicitly stacks abstraction layers, each one discarding unpredictable detail from below and predicting at a coarser categorical grain above. Talisman's claim — and it is original, not a standard observation — is that this architecture is a taxonomy. The world model research community independently reconstructed what library science formalized in the 1970s: hierarchical categorization is what intelligent systems are constrained to do when they must act over a domain. H-JEPA is not implementing a taxonomy by design; it is discovering one by necessity.

The third move is governance: AMA added a CPT taxonomy for AI billing in 2025 — assistive, augmentative, and autonomous tiers — because "AI alone cannot be billed; a service has to be classified before it can be priced." NIST published two AI taxonomies for standards application. The MIT AI Risk Repository organizes 1,725 risks from 74 frameworks. The pattern across all three bodies: taxonomy is the prerequisite for measurement, which is the prerequisite for governance. You cannot count what you have not named, cannot regulate what you have not classified.

What Talisman does not fully resolve — and it is worth naming — is the tension between hierarchical and thematic organization. Collins and Loftus, in the same cognitive science canon she cites, argued that spreading activation and co-occurrence relations are as important as hierarchy for human memory. A knowledge graph with typed edges handles both without privileging one. Talisman acknowledges this and then retreats somewhat, arguing that "the thematic requires the taxonomic" — that thematic relations only make sense once concepts are bounded. The argument is defensible. It is not airtight.

2. What the loop engineers are actually building toward

The discourse crystallized in June 2026 around a single tweet by Peter Steinberger: the job, now, is to design loops that prompt your agents. Boris Cherny replied with his own progression — from hand-coded autocomplete, to ten parallel Claude sessions, to loops that prompt agents while a few hundred other agents read his GitHub and Slack and decide what to build next. 259 pull requests in 30 days, all written by Claude Code, IDE deleted since November 2025.

What gets lost in the "loops!" excitement is the quieter observation that practitioners already figured out is the load-bearing part. Matt van Horn, synthesizing the discourse, landed on it precisely: the loop is plumbing. The asset is the skill it calls. A loop that calls a sharp, named, reusable skill compounds. A loop that re-derives what to do on each iteration just burns money — and Uber's experience confirms the denominator: they burned their entire annual AI budget in four months before the governance caught up, at a cap of $1,500 per person per tool per month. The expensive resource is no longer tokens. It is loop management, and the thing that makes loops manageable is having something stable to route to.

Thariq Shihipar at Anthropic published the formal structure in "A Harness for Every Task: Dynamic Workflows in Claude Code." The six orchestration patterns he names — classify-and-act, fan-out-and-synthesize, adversarial verification, generate-and-filter, tournament, loop-until-done — are worth reading carefully. The first one, classify-and-act, is exactly eBay's taxonomy traversal pattern expressed as an agent workflow: you classify the incoming task, then route to the handler for that classification. The loop is a wrapper around a taxonomy traversal. The fact that the taxonomy is dynamic and the traversal is LLM-mediated rather than rule-based does not change the structural identity.

Shihipar also names the three failure modes that make long-running single-context tasks unreliable: agentic laziness (stopping at partial completion), self-preferential bias (cannot objectively verify its own outputs), and goal drift (lossy compaction across turns drops edge-case constraints that were in the original prompt). Each of these is a symptom of not having explicit external structure to be grounded against. Goal drift is the most interesting: the agent re-derives what the goal is at each turn, without access to the typed, queryable definition that would let it check whether it has drifted. This is the same problem as an agent querying a database without a semantic layer — it is re-deriving the meaning of every column from context, instead of reading the governed definition that was written once and is always available.

3. Why Prukalpa is right about the substrate, and what she undersells

Sankar's "What an Enterprise Context Layer Actually Is" is vendor positioning, written by the co-founder of Atlan, and the bias should not be ignored. The five-capability taxonomy — mining, lifecycle, compounding learning, activation and retrieval, governance — maps conveniently onto Atlan's product surface. But the intellectual content survives the conflict of interest on two specific fronts.

The skills-as-functions analogy is the sharpest formulation in recent writing on this problem. Before functions, logic was copied between programs and subtly different every place it ran. Enterprise procedural knowledge is currently in that pre-function state: every time someone starts a new analysis, they re-derive what "qualified lead" means, what the correct denominator for the approval rate is, which date field to use for a given type of query. A skill — a reusable, versioned, callable unit of procedural knowledge — does for that problem what functions did for repeated logic. You write it once, with the edge cases and the caveats and the failure modes, and every downstream consumer gets it for free.

The memory-type ownership taxonomy is equally useful. Working memory and episodic memory belong close to the agent harness. Semantic memory and procedural memory belong inside the context layer. This is not Atlan's taxonomy; it is a standard cognitive psychology taxonomy applied to agent architecture. Its value is that it provides a principled answer to "what goes where" — the question that most context-layer architectural discussions either dodge or answer by accumulation.

What Sankar undersells — conspicuously — is where the bottleneck actually lives. Heqing Huang's practitioner essay "AI and Analytics, from First Principles" says it plainly: the foundation must be built by humans. The agent "reasons on top of well-defined building blocks instead of guessing at ambiguous raw tables." But the well-defined building blocks require a human conversation that often never happened, a finance decision that was never encoded anywhere but one person's memory, an analyst-PM negotiation that ended with a handshake and no documentation. ktx can ingest your Notion workspace. It cannot ingest a conversation that was never written down.

The context-layer tooling wave is moving at startup speed. The organizational wave — the decision to actually write down why a metric is defined the way it is, why that join was chosen, who owns the definition — has been moving at enterprise speed for decades. The tools are not the constraint. They never were. But having excellent tools does make the organizational debt more visible, not less: when you can see how much of the context layer is empty, the gap between "we have a context layer" and "we have populated it correctly" becomes exact and measurable rather than vague and avoidable.

4. The format hierarchy is the practitioner's decision tree

Talisman's most immediately useful contribution is a decision table she presents almost as an aside. Four formats for encoding a taxonomy, ordered by capability:

Markdown — readable but uncomputable. An agent can read it the way it reads a Slack message. It cannot query it, validate it, or check whether two entries conflict.

YAML — structured and configurable, no global identifiers. Two teams can maintain overlapping taxonomies in YAML and have no mechanism to determine whether their customer_id refers to the same concept. This is where dbt Semantic Layer, Cube, and most practical semantic infrastructure currently lives. The absence of globally addressable identifiers is not a theoretical problem; it is why cross-team metric alignment consistently requires manual reconciliation.

Spreadsheet — structured training data, one converter step from SKOS. Better than YAML for merge operations; weaker than SKOS for machine consumption.

SKOS RDF — globally addressable via URI, queryable via SPARQL, cross-scheme mappable via exactMatch and closeMatch, evaluable against stable identifiers. The format the MeSH biomedical taxonomy uses to achieve 12.5% improvement in query expansion. The format that survives organizational mergers without requiring a reconciliation sprint.

The practical guidance is not "use SKOS everywhere." The guidance is to know which limitation you are accepting when you choose YAML. If you need cross-team composability now, or federation across organizational boundaries, YAML's identifier gap costs you. If you are building inside a single organization with a single dbt project and a single metric registry, YAML plus a well-maintained schema is tractable. The choice is explicit now, rather than something you discover when the merger happens.

5. The thing to build

The convergence claim comes with a practical implication that is worth stating bluntly. If library science, world model research, and agentic engineering all needed the same missing layer and each one independently reconstructed it, then building that layer well — once, correctly, in a form that can be queried and updated without retraining — is not a vertical-specific infrastructure problem. It is a horizontal primitive that every domain that builds AI systems over structured data will eventually need.

For data infrastructure specifically, the implication is that the semantic layer is not just a metric delivery system. It is the candidate location for the explicit knowledge organization that everything else is trying to approximate. The ktx project's two-surface architecture — structured YAML for executable definitions, Markdown wiki for interpretable context, linked by typed graph edges, both git-committed and diff-reviewable — is the most principled implementation of this currently available in open source. Its gaps (no calibration layer, no coverage measurement, no refusal conditions) are exactly the gaps the previous article in this series identified. But the architecture is right: two surfaces, one executable, one interpretable, linked by typed references that stay live as the data stack changes.

What Talisman adds to this is the format prescription: the executable surface should aspire toward the SKOS end of the hierarchy, even if it starts at YAML. Not because SKOS is theoretically correct, but because global addressability is the property that lets two teams' semantic layers be compared, merged, and governed as a single asset rather than maintained in parallel as two drifting versions of the same thing.

And what the loop engineers add is the use-case: a skill library is not a nice-to-have feature of an agentic system. It is the thing that separates a loop that compounds from a loop that just runs. The semantic layer — typed, queryable, with a clean surface for agents to navigate rather than guess at — is the data analyst's skill library. Every time someone has to explain to an agent what "qualified lead" means, the skill library is missing a skill. Every time an agent applies the wrong date filter because the right one was in someone's head, the semantic layer is missing a typed definition. The convergence points at the same construction project from three different entry gates.