Why
An agent that can sequence, branch and loop over tool calls is Turing-complete, and Turing completeness is exactly what makes it unpredictable — nothing in the architecture stops an infinite loop, a conversation drifting away from its original goal, or a chain of calls running up unbounded token cost (12:46, 13:47). Pydantic-style type validation at the input catches malformed calls, not wrong ones: a well-typed call can still refund the wrong person, refund the same order twice, or set a status that does not exist in the domain. That gap between well-typed and correct is what a rule-based domain model closes, and it has to close it before the call has a side effect, not after — an agent that writes to a database or moves money does not get the luxury of a post-hoc review.
How it works
Two lineages, one guardrail: agents, ontologies and neuro-symbolic AI
Agents and ontologies are old, separate ideas that this talk pairs up. Agents go back to McCarthy and Minsky: systems that perceive, decide and act (02:33, 03:06). Ontologies go back to knowledge representation and Tom Gruber's 1993 definition, "a formal specification of a shared conceptualization" (03:42) — a structure of the entities, relationships and attributes an organization or domain actually uses (04:02, 05:30). Neuro-symbolic AI is the marriage: a neural network that can hallucinate, paired with a rule-based knowledge graph or ontology that constrains what it is allowed to conclude or do (04:12, 04:54). The ontology is not a nice-to-have documentation artifact here — it's the part of the system that is supposed to be right when the model is only supposed to be probably right.
Building an ontology: top-down, bottom-up, standards
The talk gives three ways to get one. Top-down: domain experts define the entities (an Order, a Customer) and their relationships directly, which is the same move 1980s expert systems made (06:25). Bottom-up: mine entities and relationships out of operational data, such as logs of customer interactions, rather than asking anyone to write them down (08:01). Or borrow: reuse an existing standard vocabulary — Schema.org, FOAF, DBpedia — instead of inventing one (08:16, 08:55). In practice these combine: standards for the common vocabulary, top-down for the domain-specific core, bottom-up to keep it honest against what the system actually sees.
RDFS/OWL reasoning: domain and range, transitive, functional properties
Once entities and relationships exist, RDFS and OWL add inference and validation on top of the graph (09:24, 12:07). Domain and range let the reasoner work backward from a single fact: told only "Bob teaches Scooter," it infers Bob is a teacher (and a person) and Scooter is a student, because "teaches" is declared to run from teachers to students (10:11). Transitive properties propagate a relationship along a chain — an ancestor relation from A to B and B to C implies A to C (10:49). Functional properties assert a relationship must resolve to exactly one value; the example given is a biological father, which lets the reasoner flag a contradiction or recognize that two records describe the same entity when the constraint would otherwise be violated (11:23). None of this is exotic: it's the same domain/range, transitive and functional building blocks description logic has always had, applied to whatever an agent is about to do.
The agent loop's three risks, and two gates: door and ledger
Giving an agent the ability to sequence, branch on conditions and loop over tool calls makes it Turing-complete — and buys three specific risks along with the power: infinite loops, drift away from the original context mid-conversation, and unbounded token cost (12:46, 13:47). The talk's fix is architectural, not a better prompt. An LLM never executes a tool directly; it only ever generates the parameters for a call, so something has to sit in the middle and mediate every call (15:16, 16:08). That middle layer gets two gates. At the door, Pydantic validates the tool call's input types strictly — is this a number, is that a string, before anything runs (18:14, 18:23). At the ledger, before a tool's result or an agent's decision is allowed to change domain state, an ontology reasoner or validator checks it for logical consistency (16:51, 18:23). The door catches malformed requests; the ledger catches well-formed ones that are still wrong.
Three refunds a prompt would miss
The talk grounds all of this in one domain — refunds — because natural-language instructions fail there in specific, repeatable ways that OWL constraints catch (19:11, 19:53):
- A duplicate refund on the same order, blocked by a uniqueness constraint (19:11).
- A refund routed to a support rep instead of the buyer, blocked by declaring
CustomerandSupport Repdisjoint classes (19:26). - An invented status value, blocked by declaring status an enumeration —
paid,shipped,refundedand nothing else, so a hallucinated "probably shipped" has nowhere to land (19:34).
Coyle's framing, per the summary: an agent that is going to be trusted with real side effects — a database update, a financial transaction — needs this symbolic validation layer, because the agent itself is only ever probabilistic (18:23, 20:03).
Where it lands in Jayverse
- Rabbit: the EIP-7715 mandate enforcers are the ledger-side validator. A session-key mandate's cap, expiry and allowed-target list are already OWL-like constraints written in Solidity instead of RDFS; a disjoint-class rule such as "payer ≠ payee's support agent" has a direct analogue in "the address you're paying is not an address you also operate."
- Verex: market status is an enumeration, resolution is a functional property. A market's status field should hold only one of a fixed set of values, never a hallucinated in-between state, and resolution is functional by construction — one outcome per market, and a second write should fail the same way a functional-property violation would.
- Auditor: "by which rule" is the ontology. The rules the Auditor checks against belong in a schema or a graph, not in a paragraph of prose — store them as data so the check is something that runs, not something that's read.
- Theory: domain/range, transitive and functional properties are description logic. Worth its own entry next to the scheduling notes — these three constraint types are the textbook building blocks of description logic, and the refund examples are a working illustration of each.
- Dark Horse: the ontology is the machine-checkable half of a boundary (e) file. A boundary doc says what's out of bounds in prose; an ontology-shaped validator is the part of that boundary that a call actually has to pass before it executes.
Verified and unverified
Verified on 2026-09-19: Tom Gruber's 1993 definition of an ontology ("a formal specification of a shared conceptualization") is a real, widely cited definition; RDFS and OWL are W3C standards that define domain/range inference, transitive properties, functional properties and disjoint classes; Schema.org, FOAF and DBpedia exist as published vocabularies and knowledge bases; Pydantic validates Python types at runtime; "neuro-symbolic AI" is an established term for combining neural networks with rule-based or symbolic systems. Taken from the summary and not independently checked: the talk's existence, exact length and venue beyond the linked video, the attribution to Frank Coyle as a UC Berkeley professor (stated here per the summary only, no further biography checked), every timestamp, the worked examples (the teaching, ancestor and biological-father illustrations, the three refund scenarios), and the closing line "Nothing is a mistake, only make," quoted as given.
Sources: YouTube — Why Agentic Systems Need Ontologies (AI Engineer conference talk, per the summary) · related items: Tech #62 (agentic engineering writes the boundaries), Tech #102 (mlflow-tracing-llm-as-judge — evaluation after the fact vs. validation before the side effect), Tech (pocock-fundamentals-matter-more).
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| ontology | 온톨로지(도메인의 엔티티·관계·속성을 formal하게 정의한 구조) · 이 항목 전체의 핵심 개념. "a formal specification of a shared conceptualization" |
| neuro-symbolic AI | 뉴로-심볼릭 AI(신경망과 규칙 기반 시스템의 결합) · 온톨로지가 신경망의 가드레일로 쓰이는 접근 전체를 가리킴. "a neuro-symbolic layer next to the neural network" |
| formal specification | 형식적 명세(모호함 없이 규칙으로 적은 정의) · 그루버의 온톨로지 정의에서. "a formal specification of a shared conceptualization" |
| top-down / bottom-up | 하향식 / 상향식(전문가가 정의 대 데이터에서 추출) · 온톨로지를 만드는 두 반대 방법. "mine entities and relationships out of operational data" |
| domain and range | 정의역과 공역(관계가 어떤 타입에서 어떤 타입으로 가는지 제약) · 한 문장만으로 역추론하게 해주는 OWL 개념. "Domain and range let the reasoner work backward from a single fact" |
| transitive property | 이행 속성(A→B, B→C면 A→C가 성립) · 조상 관계 같은 연쇄 추론에 쓰임. "Transitive properties propagate a relationship along a chain" |
| functional property | 함수적 속성(관계가 정확히 하나의 값으로만 귀결) · 모순 검출·동일 개체 식별에 쓰임. "Functional properties assert a relationship must resolve to exactly one value" |
| disjoint class | 서로소 클래스(두 클래스가 절대 겹치지 않는다는 선언) · 환불이 상담원에게 가는 걸 막는 제약. "declaring Customer and Support Rep disjoint classes" |
| Turing-complete | 튜링 완전(순차·조건·반복을 가지면 이론상 무엇이든 계산 가능) · 에이전트가 예측 불가능해지는 이유로 언급됨. "An agent that can sequence, branch and loop over tool calls is Turing-complete" |
| drift | (맥락) 탈선·표류 · 대화가 원래 목표에서 벗어나는 현상. "a conversation drifting away from its original goal" |
| hallucination | 환각(모델이 사실이 아닌 것을 그럴듯하게 생성) · 신경망에 심볼릭 검증이 필요한 이유. "a neural network that can hallucinate" |
| guardrail | 가드레일(범위를 벗어나지 못하게 막는 안전장치) · 온톨로지가 신경망에 대해 하는 역할. "Two lineages, one guardrail" |
| side effect | 부작용(호출이 도메인 상태를 실제로 바꾸는 효과) · 검증이 반드시 이보다 먼저 끝나야 한다는 맥락. "before the call has a side effect" |
| enumeration (enum) | 열거형(정해진 값 목록만 허용) · 상태값 환각을 막는 제약. "declaring status an enumeration" |
| description logic | 기술 논리(온톨로지 추론의 형식 논리 기반) · domain/range, 이행, 함수적 속성이 여기서 나온 개념. "description logic has always had" |
| door / ledger | 입구 / 원장(이 항목이 쓰는 두 단계 검증 지점의 비유) · 입구=타입 검증, 원장=의미 검증. "validate a tool call's types at the "door" with Pydantic" |
| RDFS | Resource Description Framework Schema(자원 기술 프레임워크 스키마) · 그래프에 추론을 더하는 W3C 표준 중 하나. "RDFS and OWL add inference and validation on top of the graph" |
| OWL | Web Ontology Language(웹 온톨로지 언어) · domain/range, 이행·함수적 속성, disjoint 클래스를 표현하는 W3C 표준. "repeatable ways that OWL constraints catch" |
| LLM | Large Language Model(거대 언어 모델) · 확률적이고 탈선하기 쉬운, 이 항목이 검증 계층을 요구하는 대상. "LLM-driven agents are probabilistic and prone to drift" |
| Pydantic | 파이썬 런타임 타입 검증 라이브러리 · "입구" 단계에서 도구 호출의 타입을 검증하는 도구. "Pydantic validates the tool call's input types strictly" |