Sebastian Schiller-Stoff

Practice note AI agents

EN — current language DE — Deutsch

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.

Practice note · record 2025–2026

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.

  1. Brief

    Names the task, the systems in scope, the tools and the stop conditions.

  2. Act

    Uses only the listed tools, one step at a time.

  3. Check

    Passes the written test before anything is kept.

  4. 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.

  1. It runs

    The task completes on a real input, not a demo.

  2. It is tested

    A test proves the change, and that test is in the repository.

  3. It is attributable

    The prompt, the sources and the diff are recorded.

  4. It is licensed

    Every dependency and its terms are known.

  5. 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.

  1. Presentation

    Wo ist der KI-Sweetspot?
  2. Conference paper

    Wo ist der KI-Sweetspot? Nutzen und Herausforderungen für die Einbindung von KI-Assistenzsystemen ins Geisteswissenschaftliche Asset Management System (GAMS)
  3. Conference abstract

    Kann LLM generierter Code FAIR sein? Herausforderungen für das Forschungsdatenmanagement
  4. Workshop slides

    Introduction to AI-Agent-Engineering
  5. Workshop

    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.

Colophon

Method
A practice note, 14 November 2026. It describes the boundary I use in agent work and teach in workshops; it is not a specification.
Sources
The five records listed in chapter 04: the 2025 agent-engineering workshop and its slides, the 2026 GAMS paper and talk, and the 2025 FAIR-code abstract. All are open on Zenodo, and every claim in this note maps to one of them. Its longer reading is Agents are roles, not robots.
Record titles
Titles follow the deposits. Normalised: typographic quotation, terminal punctuation, and the GAMS paper’s deposited „Assiszentzsystemen“, corrected above. Type labels name the event form; the deposits’ own types are broader.
Authorship
Sebastian David Schiller-Stoff. The work in the record was co-authored with Münzer, Scholger, Pollin, Steyer and Citro. This is a practice note, not a product claim: it describes a method I use and teach, not a standard or a certification.
JavaScript
None. The page is text, CSS and a small inline sprite.
Licence
Text and layout: CC BY-NC 4.0. The cited records keep their own Zenodo licences.

Back to the CV AI coding assistants