How to create and use an ontology with our AI-Assisted Ontology Modelling Methodology, and how an agent harness needs to be implemented to govern agents in production.
In the first two parts of this series, I described the Five Agent Failure Patterns and the fixes – a knowledge graph built on an ontology, with an agent harness governing how agents use them. This article talks about how one of the most important pieces, the ontology, gets built and, of equal importance, how the agent harness is used to govern the agents.
The stakes are real. Even the experienced teams we see are not getting the accuracy they need, typically reaching only 80% with their AI agents, and some organisations are below that. That is not accurate enough for production, where enterprises need 90–100% accuracy.
Speed without accuracy isn't progress – it's exposure.
Everything an agent does downstream depends on the ontology. If the definitions are wrong, the knowledge graph is wrong, the rules are wrong and the agent is confidently wrong.
This is also where many organisations stall. Skilled ontology engineers are scarce, and modelling by hand can take months. Many teams take on the ontology themselves, and this is where agent accuracy is most often won or lost – the quality of the ontology sets the ceiling on how accurate the agents can be.
Handing the job to an LLM on its own is fast, but it introduces structural defects that compound as the model grows: over-deep hierarchies, inconsistent naming, duplicated properties, and things modelled as types when they are really individual items. The first few classes look reasonable. The problems appear later, once other systems depend on them. This is why Graph Research Labs designed the AI-Assisted Ontology Modelling Methodology – to help organisations build better ontologies and get them into production.
The AI-Assisted Ontology Modelling Methodology, which Graph Research Labs presented at the 2026 Knowledge Graph Conference in New York, combines the two but does not replace the expert ontologist. Instead it speeds up the process by using an LLM and a set of specialist AI agents to do the legwork, while the ontologist stays in the loop at every decision point and makes the design judgements an LLM cannot. It has seven stages.
An ontology engineer reviews the initial version of the ontology, whether it was adapted from an industry standard, drafted by the organisation's own team or generated with an LLM. The review sets the baseline: what is sound, what is missing and what must not change.
An LLM generates competency questions – the questions the ontology must be able to answer for the agents that will rely on it. These become the tests. If the ontology cannot answer them, an agent working from it will not be able to either.
AI agents are trained and deployed to critique the ontology, each looking for a different kind of defect. They cover ground far faster than a person could, and they do it the same way every time.
The engineer assesses the agents' critique, decides what to change and applies the fixes. This is the heart of the methodology: the agents find and propose, and the expert decides. Design judgement stays with a person.
The changes are validated with validation tooling, by both the engineer and the LLM, to confirm that the fixes hold and that nothing else has broken.
An LLM produces a conformance report that scores the changes to the ontology. A poor score sends it back for further fixes.
The ontology moves into governed production, where the knowledge graph and the agents begin to depend on it.
Any stage can be re-entered as the ontology evolves.
Stages 6 and 7 are where the model is proven. Using GRL Generators, an ontology can be production tested with a full data flow in hours, and the cycle repeated until it reaches production level for release. Testing the model against real data before release, rather than after, is what saves months – and in some cases years – of effort.
Agents appear twice in this picture. They help build the ontology, in the critique stage, and they are what the ontology ultimately serves once it is released.
The same principle applies to both. An agent left to invent structure produces fluent, confident and unreliable results. An agent that works within a well-thought-out structure that a person has defined accurately becomes a dependable assistant.
With a production-ready ontology in place, two steps connect it to the organisation.
Scope by one decision. Choose a single decision an agent should support, and let its competency questions set the boundary. The ontology starts small and grows with each new use case.
Connect the data. Map the relevant data from each IT system into the knowledge graph using the ontology's definitions. Anything that does not fit is flagged for review rather than forced in, because a gap in the model is a signal that something is missing.
We now introduce the agent harness in more detail, as the governance layer that sits around an agent. With the ontology and the knowledge graph in place, we add the agent harness to control the AI agents and to make it all work more accurately.
It supplies what the agent needs. For each task, the harness gives the agent the relevant part of the ontology and the connected facts from the knowledge graph. The agent starts from the organisation's own definitions rather than searching for something that looks similar.
It checks every action before it runs. Business rules move out of the prompt and become constraints held alongside the ontology. The agent proposes an action, the harness checks it against those rules, and an action that breaks a rule is blocked before it executes, with the rule cited. This is governance by structure, not by chance.
It governs what agents remember. Anything an agent writes back or stores for later passes through the same checks. One agent's mistake does not quietly become another agent's fact.
It records why. Each decision is captured with the facts it used, the rule it was checked against and the version of the data it saw. When an auditor, regulator or customer asks why the agent did something, the answer is a query rather than an investigation.
The agent still does the reasoning. However it can only act on what the organisation has defined as valid. And because the harness reads the ontology rather than hard-coding the business, the same harness can govern a compliance agent and a clinical agent – only the ontology differs.
Getting to production is only half the job. Organisations change – new products, new regulations, restructures – and an agent that was accurate in March can quietly drift by September.
This is where the approach pays off a second time. Because the meaning and the rules live in the ontology rather than in each agent, a change is made once and every agent picks it up. The methodology is re-entered at whichever stage the change calls for, so the ontology improves with use rather than decaying.
The organisations that win the AI race will be those that move quickly while keeping their focus on accuracy. There is little point in marketing AI excellence if, under the covers, the agents never became accurate enough for the organisation or its customers to rely on. The practical way to get there is to start with a single decision an agent will support and take it through a full test life cycle:
Once that first agent is running accurately, the same ontology, knowledge graph and harness are extended for the next decision rather than rebuilt. Each new agent starts from what the organisation has already defined – which is how a single pilot becomes a set of agents an organisation can trust.
A seven-stage method developed by Dougal Watt of Graph Research Labs and presented at the 2026 Knowledge Graph Conference in New York, in which expert ontology engineers build production ontologies with the help of an LLM and specialist AI agents, staying in the loop at every decision point: Review, Define, Critique, Assess, Validate, Conform and Release.
It can produce one quickly, but without constraints it introduces structural defects that compound as the ontology grows. LLMs and AI agents are most useful when an expert ontology engineer defines the structure and makes the design decisions.
An agent harness is a governance layer around an AI agent. It combines an ontology, a knowledge graph, automated rule checks and a record of every decision, so the agent can only act on what the organisation has defined as valid.