Practice note · AI agent engineering
Agents with a brief.
An agent is not a colleague. It is a bounded worker: a written brief, a set of tools, a check it must pass, and a record of what it did. I design that boundary into research systems — and teach it.
The brief
An agent is a loop: read, choose a tool, act, check, repeat.
The model is one component; the brief decides whether the loop is useful. A brief names the task, the systems in scope, the tools, the stop conditions and the signatory.
Research software is full of actions that are easy to automate and expensive to undo: a migration, a bulk metadata edit, a publish step, a deletion. The brief is how they stay reversible.
-
Brief
Names the task, the systems in scope, the tools and the stop conditions.
-
Act
Uses only the listed tools, one step at a time.
-
Check
Passes the written test before anything is kept.
-
Record
Leaves a trace of what it read, ran and changed.
The loop is deliberately dull. Dull is the point.
The toolbox
The useful tools are unglamorous: an archive API, a search index, a repository checkout, a test suite, a metadata schema.
The 2026 DHd paper on AI assistance in GAMS asks where the sweet spot is between reusing existing LLM building blocks and building your own — and how to integrate either without technical debt. The lesson I take is narrower than the demo: give the agent the same interfaces a colleague would get, and no more.
In scope
- Read public or licensed sources
- Query the archive API
- Draft code and tests
- Propose metadata edits
- Summarise a diff
Out of scope
- Publish or delete records
- Invent citations or identifiers
- Change licences
- Act on personal data
- Approve its own output
A callable tool is both an attack surface and an audit obligation. Keep the list short.
The check
The 2025 FAIR-code abstract with Münzer asks whether LLM-generated code can be FAIR.
Not by default: generated code often arrives without maintainable structure, metadata or a clear licence, which limits findability and lawful reuse. The same is true of an agent’s output, only faster.
So the check is a gate, not a vibe.
-
It runs
The task completes on a real input, not a demo.
-
It is tested
A test proves the change, and that test is in the repository.
-
It is attributable
The prompt, the sources and the diff are recorded.
-
It is licensed
Every dependency and its terms are known.
-
A person signs it
The commit carries a human author, not the tool.
What I teach
The method is public; the record is the evidence.
Agent engineering is taught as a bounded practice, not a tool tour.
- Wo ist der KI-Sweetspot?
- Wo ist der KI-Sweetspot? Nutzen und Herausforderungen für die Einbindung von KI-Assistenzsystemen ins Geisteswissenschaftliche Asset Management System (GAMS)
- Kann LLM generierter Code FAIR sein? Herausforderungen für das Forschungsdatenmanagement
- Introduction to AI-Agent-Engineering
- Fortgeschrittenes Prompt und AI Agent Engineering für wissenschaftliche Textproduktion
The workshops and papers behind this record were co-authored with Münzer, Scholger, Pollin, Steyer and Citro.