Workshop summary: Agentic Work with University LLMs

28 October 2026 · DH Lab + Zoom · led by Prof. Dr. Achim Rabus · about 40 participants in total (on site and online)

Key claims

1. From chatbot to agent. An LLM used as an agent has a different job than a chat window: it gets things done. It uses tools (file system, terminal, web, computer vision), receives a goal rather than a single question, and runs in loops – it fails, corrects itself, and reaches the goal. That closes the gap between what we want done and what we are able to produce ourselves.
2. The harness matters more than the newest model. A good agent harness (OpenCode, Hermes) plus a sufficiently good model was described as more productive than the latest frontier model in a chat window. The harness controls tools, loops, workflows, and skills.
3. Hallucinations become manageable in the agentic mode. LLMs predict the next token and have no concept of right and wrong – as a chatbot, that is a serious problem. In the agentic mode the work runs through code, and code either runs or it does not: a wrong bracket or variable makes the script fail, the agent sees it and tries the next path.
4. Natural language as a programming language. Agentic work amounts to natural language programming – and that is a discipline in which philologists are well placed: precision in language use becomes a productive force.
5. Data sovereignty. Commercial frontier services (GPT, Claude, Gemini) are capable, but poorly suited for administrative contexts and sensitive data: unclear data flows, metadata, options that may be re-enabled after an update, dependency, and quota limits in the middle of work. The university provides open models through OpenWebUI, reachable by API key from the university network.

Live demonstrations

• Renaming ISBN PDFs (OpenCode, ufr/coding-complex): a folder of PDFs named after ISBNs; the agent researched the metadata, asked which naming scheme was wanted (author, year, title) and produced renamed copies with diacritics preserved. The task took over ten minutes and roughly 40,000 tokens, which is quite a lot.
• Student statistics → dashboard (Hermes): from freely available statistics PDFs of the university website (faculties, student numbers, several years), the agent built an interactive time-series visualisation as a single HTML file.
• Dictating instead of typing: a local open-source speech-to-text tool (handy.computer) was used instead of a cloud service with unclear data handling; prompts in German worked just as well.
• Controlling reasoning levels (Hermes): the thinking level is adjustable per model; "medium" is a good starting point – extra high or max burns tokens very quickly without a significant gain.
• The agent as system administrator (anecdote): before the workshop, an agent set up an SSH tunnel to a workstation on its own and configured the LLM interface there; after several failed attempts it improved its method by itself. Earlier in the year, one agent deleted a file without being asked to, and later sent an email to the Pretix support on its own initiative – lessons on blast radius and the need to watch what agents do.

Technical findings

• Context window: the default context window is about 256k tokens; the university instance used in the workshop was reported to offer 1M – a whole working session fits into the window. At 90 % fill level, auto-compact kicks in: the agent condenses the interaction, and quality decreases in the process.
• Model load: when many users pull the model at once, the university instance can become slow or stall – workaround: switch to another model (ufr/coding-fast, gemma-4-31b-llmlb) and retry later.
• Aliases: ufr/coding-complex and similar aliases abstract away the underlying model (the computing centre can update it transparently) – convenient for everyday work, useless for model comparisons.
• Safety: a dedicated working folder, copies rather than originals, confirmations turned on, and knowing the limits of the folder sandbox – for higher demands, a Docker sandbox or a VM (trade-off: considerable installation effort).

Best practices (wrap-up)

• Plan first, then run: for complex tasks, have the agent write a plan into a Markdown file and check that plan before letting it run – avoid one-shot prompting.
• Control sycophancy: ask the agent whether it has everything it needs, and push it to disagree honestly – pleasing answers help neither the project nor the human.
• Work in parallel: several agents on several tasks at the same time; waiting time is time for other work.
• Avoid over-engineering: a small dataset and a simple visualisation call for a single HTML file rather than a database and a framework – easy to send around and to upload to ILIAS.
• Backups: two independent, traditional routes (university cloud or network storage plus an external drive) – set up by yourself, with the agent kept out of it. Data loss so far came from hardware failure, and never from agents.
• APIs as data sources: agents are good at working out how to talk to website databases and platforms – try it out, and stay in charge.
• Limit cognitive offloading: methods that you still need to master yourself should be kept in your own hands.
• Skills: GrillMe (hard questions about what you actually want, instead of pleasing answers), Learning Opportunities, and Superpowers – all essentially text files that the agent can use, and they noticeably raise the quality of collaboration.
• Git and GitLab Pages: version control for non-programmers as well; with GitLab Pages you can publish a website with a fixed URL on your own, without depending on the University computing center.

From the Q&A round (not in the video)

• Python is the default language of the agents; other languages (R for statistics, for instance) have to be requested explicitly.
• For diacritics-sensitive text work (a Greek text), the agent chose the deterministic route (a Python script rather than world knowledge); when results deviate, make the prompt more precise and name the tool explicitly.
• Installation hurdles (npm post-install scripts, PowerShell execution policy) were solved together during the workshop and have gone into the setup guides.

Take-home

Agentic work moves the emphasis from executing to planning, structuring, and delegating – with more thinking, and more. For the humanities this is an opportunity: precise language and argumentation are the core competences of this way of working.