Richard Scheiwe

Building AI systems. Writing about how AI is changing work, human judgment, and intelligence. 10+ years engineering & leadership.

[Richard Scheiwe @richardscheiwe]

Most agent evaluation measures whether the output looked right. The harder question is whether the agent was reasoning against anything stable. A system with no represented domain can pass your eval on Monday and fail the same task on Tuesday, because there was never a fixed world underneath the fluency.

[Richard Scheiwe @richardscheiwe]

Every scheduling system eventually rediscovers the same thing: a booking has a lifecycle, not just a status field. Requested, confirmed, checked in, completed, no-showed, and each transition has rules. Model the lifecycle as knowledge and every service agrees on what can happen next, instead of each one guessing separately.

[Richard Scheiwe @richardscheiwe]

A taxonomy tells you a cat is a mammal. An ontology tells you what being a mammal commits you to: it breathes, it regulates its own temperature, it cannot also be a species of fish. The first is a label you attach. The second carries consequences a system can actually check.

[Richard Scheiwe @richardscheiwe]

Bigger context windows keep getting sold as the fix for reasoning, but a longer window only lets the model see more unstructured text at once. More to read is not more to reason with. What changes the answer is whether the facts in that window are typed and related, or just sitting next to each other.

[Richard Scheiwe @richardscheiwe]

Retrieval gives an agent the right paragraph. It does not give the agent a world. You can pull the passage that says an invoice is overdue and still have no system that knows an overdue invoice blocks the next shipment. The paragraph is evidence. The rule is knowledge, and it has to live somewhere the agent can reason over rather than just read.

[Richard Scheiwe @richardscheiwe]

The most common reaction to talking about knowledge representation is not disagreement. It is recognition. People have hit the wall where prompts and embeddings stop being enough, and they did not have a name for what was missing. The name is structure: typed, durable, queryable knowledge.

[Richard Scheiwe @richardscheiwe]

Aristotle was sorting the world into categories with defining properties two thousand years ago. OWL classes do the same job with more rigor. The formal study of what exists turns out to be remarkably stable. What changed is that machines now need the answer in a form they can compute over.

[Richard Scheiwe @richardscheiwe]

A constraint like an account cannot be both closed and active is not a database rule and not a validation check. It is a fact about the world your system lives in. Put it where the meaning is, and every part of the system inherits it for free.

[Richard Scheiwe @richardscheiwe]

The first time you watch a reasoner derive a fact nobody wrote down, knowledge representation stops being abstract. You said managers are employees and that this person is a manager. You never stated they are an employee. The system concludes it. Now scale that to thousands of facts.

[Richard Scheiwe @richardscheiwe]

Most make-the-agent-more-reliable work is really give-the-agent-something-stable-to-be-reliable-about. A model with no fixed world to check against will be fluent and confidently wrong. Reliability is downstream of having a represented domain, not a better prompt.

[Richard Scheiwe @richardscheiwe]

The semantic web was early, not wrong. RDF and OWL outlived the hype and quietly run real infrastructure in healthcare, life sciences, and enterprise integration. The vision arrived before the systems that needed it. Those systems are here now, and we call them agents.

[Richard Scheiwe @richardscheiwe]

Orders, vendors, shipments. The bugs live in the relationships: an order with two vendors, a shipment dated before its order, a quantity that should never be negative. Cardinality and temporal constraints at the knowledge layer stop bad data before it becomes a bad decision downstream.

[Richard Scheiwe @richardscheiwe]

Research note: GeneralVLA-2 pairs geometry-aware reconstruction with what the authors call governed memory for robot planning.

The phrase governed memory is the tell. Once an agent acts in the world over time, raw recall is not enough; it needs memory with rules about what can be written, trusted, and overwritten. That is a knowledge-representation problem wearing a robotics hat.

https://t.co/TWRMmxmGxz

[Richard Scheiwe @richardscheiwe]

JSON tells software what shape your data has. It cannot tell a system what your data means, which is why just use JSON quietly breaks the moment you need inference. Shape is not meaning. The difference shows up exactly when the work gets hard.

[Richard Scheiwe @richardscheiwe]

OWL is twenty years old. The problems it set out to solve, meaning, constraints, inference, are more relevant now than they have ever been, because agents need something stable to reason against. The concepts held up. The syntax and tooling did not keep pace. That gap is worth closing.

[Richard Scheiwe @richardscheiwe]

Compliance is not a layer you bolt on. It is a property of the domain. A transaction that moves money has invariants: balances do not go negative, debits equal credits, some transfers require approvals. Encode those where the meaning lives and an auditor can actually read them.

[Richard Scheiwe @richardscheiwe]

Research note: MemSlides uses a hierarchical memory to let an agent revise a long artifact over many turns without losing the thread.

What stands out is the move away from flat context toward structured memory with levels. The agents that hold up over long tasks tend to be the ones whose memory has shape, where what matters is organized rather than just accumulated and re-scanned.

https://t.co/xm7vveb0Oj

[Richard Scheiwe @richardscheiwe]

A patient encounter is a lifecycle: scheduled, arrived, in progress, discharged, billed. Each transition has guards and consequences. Bury that in application code and every service reimplements it slightly differently. Put it in the knowledge layer and the whole system agrees on what is allowed to happen next.

[Richard Scheiwe @richardscheiwe]

Modeling a company sounds trivial until you try to write down what is actually true. An employee belongs to exactly one company. A manager is an employee. A department has a head who is also an employee. None of that is data shape. It is the rules the data must obey, and most schemas never state them.

[Richard Scheiwe @richardscheiwe]

The case for new ontology tooling in 2026 is simple.

The problems OWL set out to solve, meaning, constraints, inference, are more relevant than ever now that agents need to reason. The concepts held up. The syntax and tooling did not keep pace.

That gap is the whole reason to build.