Explore projects

Modelica framework for the dynamic simulation of aqueous electrolyte systems under consideration of dissociation and phase equilibria

This project contains various helpful tools and scripts for RDM.

Cricket consists of two parts: A virtualization layer for CUDA applications that allows the isolation of GPU and CPU parts by using Remote Procedure Calls and a checkpoint/restart tool for GPU kernels. Logo by Freepik.com

A tool for adding metadata to tables

JupyterHub Services

Abgabepartnerbörse

The plotID toolkit supports researchers in tracking and storing relevant data in plots. Plots are labelled with an ID and the corresponding data is stored.

Create all documents for your next CRC/SFB in an easy and efficient way.

Latex Design der AG Halfmann

A collection of tools for slip system analysis. Please cite our review paper and tutorial if you use the code provided in this project.

This repository contains the application which tries to detect change on data resources and describes them using PROV-O and metadata extraction methods.

Title: Capable AI Coding Agents

Author: Robin Rentmeister

Abstract: The use of Large Language Models (LLMs) in software engineering is shifting from single-shot code completion to autonomous, multi-step workflows that operate over large repositories and interact with external tools and APIs. While this transition promises productivity gains, it also introduces concerns for correct and secure results. In contrast to isolated snippet generation, agentic pipelines can amplify small mistakes into cascading failures, exhibit non-deterministic behavior, and produce plausible code that silently violates security or domain constraints.

This survey reviews failure modes in agentic coding, consolidates techniques for improving correctness and reproducibility, and highlights remaining challenges. We organize assurance mechanisms into three complementary layers: (i) operational infrastructure for deterministic execution, provenance, and tracking; (ii) cognitive architectures that couples generation with executable checks and iterative repair; and (iii) interaction paradigms for human and multi-agent interaction, policy enforcement, and governance. Drawing on evidence from execution-based benchmarks and safety evaluations, we summarize which mechanisms yield consistent improvements and where concerns persist. We conclude by outlining open challenges toward scalable, long-horizon robustness and verifiable reliability for AI coding agents.