Scientific Computing Core

AI models and surrogate models

Scientific Computing Core

Agentic AI for Research Software

CASUS Scientific Computing Core always looks for technology, tools and solutions to improve research software development. AI assisted coding is certainly one of the more disruptive advances. It started to be (barely) usable in the beginning of 2025 and showed tremendous improvements since then.
 

Using AI Agents for Software Development

We use AI agents for our own software development projects to constantly learn about pros and cons and to stay up to date with the state of the art. We compare ability and quality and different ways to use them for different purposes. This includes vendor LLMs as well as self-hosted LLMs (within Helmholtz or HZDR as well as locally hosted by us) to track their differences and improvements.

We try different APIs, tools, and IDEs and teach other scientists in CASUS, HZDR, and beyond about new and relevant solutions. And we pay attention to privacy and confidentiality and consequences of AI usage.

Developing Solutions for Responsible AI Coding Assistants

CASUS Scientific Computing Core is also working on Open Source solutions to improve Agentic AI coding. This includes novel features, ease of use as well as security aspects.
 
With the Terok agentic coding environment we bring a solution to manage many projects, agents, and agent tasks in a secure way.
 
AI agents pose additional challenges to security and confidentiality compared to AI chatbots. Agents can read local files and actively run commands. This is a risk in case the agent performs malicious actions. At the same time this is the huge benefit of the agentic way. Restricting this ability would give up most of the advantages.
 
Therefore, Terok puts the AI agents into isolated environments where we can benefit from their abilities without those risks. Furthermore, it manages multiple such environments, safeguarts the access to public software repositories, and allows to easily switch between different vendor LLMs or self-hosted LLMs.
 
 
Terok management view with one project (left) and one task (right)
Terok with an agent session using opencode.ai

Dr Andreas Knüpfer

Dr. Jiří Vyskočil