Selected writing · AI agent engineering
Agents are roles, not robots.
The 2025 agent-engineering workshop did not start with a framework. It started with a question: which human role does this agent stand in for? That question still decides whether an agent is useful.
Why it exists
“Agent” arrived in the Digital Humanities as a vendor word.
In early 2025 the word was doing two jobs at once: a product category and a design pattern. The workshop — a section of the advanced prompt and agent-engineering workshop at DHd 2025 in Bielefeld — tried to separate them. Its showcase was not a chatbot. It was a group of cooperating agents producing teaching material for DERLA, the digital memorial landscape.
The audience was mixed: people who wrote Python and people who commissioned it. So the workshop taught one language for both — Python, a commercial model provider, and an object-oriented model of the work.
The aim was deliberately modest: a basic understanding of agent engineering, enough to build a semi-automated system and enough to judge one.
The argument
Three words, three disciplines: AI, agent, engineering.
The AI part is the smallest. Prompting, context and evaluation are the everyday skills; training and inference are the machine room. The workshop spent its time on the difference between using a model as a knowledge base and using it as an interface to a system.
The agent part is software design. The workshop’s definition is narrow on purpose: an agent is a code abstraction around an LLM, built around a human role in a domain — data stewardship, cataloguing, research assistance — and it should mirror how that domain divides its labour. A framework is optional. Full autonomy is optional. A role is not.
The engineering part is research software engineering: versioning, tests, cost, evaluation. The demonstration was a small crew — a planning agent, a scanner, price analysts, an ensemble that compares them, and a messaging agent. Each one is legible on its own.
In the workshop’s definition
- a named role
- a domain division of labour
- explicit tools
- a written task
- an evaluation step
Not required
- an agent framework
- full autonomy
- one vendor
- a chat window
That is the whole argument. The model is a component. The role is the design.
What holds in 2026
The tools moved faster than the definition.
Two years on, the frameworks have consolidated and the interfaces have moved from chat to the change set. What did not move is the useful unit of design. An agent that cannot name the role it stands in for is a demonstration.
The 2026 GAMS paper asks the same question one level up: where the sweet spot lies between reusing model building blocks and building your own, and how to integrate either without technical debt. The workshop’s answer was to keep the boundary small.
What I would change: teach cost and evaluation first, not last. In 2025 the pricing slide came after the architecture; in 2026 the bill is part of the architecture.
And I would say the quiet part louder: a role is also an accountability. Someone signs the output. The workshop said this in its engineering section; it belongs at the front.
The record
The deposit is a lesson, not a paper: slides and four notebooks, open for reuse.