Two of TUM Research Data Hub’s Data Stewards (one at MDSI, one at UB) are part of ODSL, working at exactly the intersection of data-intensive research and everyday scientific practice where these questions come alive. On the ODSL Agentic AI Day it was impressive to see how directly the technical answers turned into research data management questions:

Documentation becomes operational
An agent decides what to do by reading descriptions: the docstrings on the functions it can call, the conventions written into an AGENTS.md file, the plain-text definitions that say how a task is done here. Documentation used to be written for the next human — the PhD student inheriting the pipeline. It is now also the interface a machine reads to decide what to do. Metadata and documentation no longer only make research findable; they deterimine what gets done to it.

Less context, better results
The instinct is to give the model everything and let it work out what matters. The benchmark shown on the day says otherwise: across several current models, accuracy fell steadily as the input grew. Well-structured, well-described, appropriately granular data isn't only good practice.

Provenance and audit trails
An agent operating with a researcher's credentials can read files, run analyses, and edit outputs. Which tool ran, on which data, producing which result — the questions research data steward has always asked — become the only way to reconstruct what happened, and the basis for keeping sensitive data out of reach in the first place.

Reliability, reproducibility, provenance, and the protection of sensitive data aren't obstacles to using agentic AI in science; they're what makes it trustworthy. Capability is arriving faster than the guardrails around it and closing that gap is work we're already set up to do — through consultation, training, and the documentation and metadata practice we support across TUM.

Thanks to the ODSL team for a genuinely useful afternoon. With colleagues working across both worlds, these are exactly the conversations we want to be part of and to help shape.

The TUM Research Data Hub supports researchers navigating these questions. 
Get in touch: researchdata@tum.de

Detailed event program: https://indico.ph.tum.de/event/8034/

More information about ODSL: ODSL / ODC | Origins-Cluster