Bridging SW Engineering, MLOps, and Operational Agents: Foundations for Agentic AI in Cyber-Physical Systems

Cyber-physical systems (CPS) integrate communication and control for real-time interaction with the physical world and increasingly use various AI agents for tasks such as optimization and adaptive control. However, seamless collaboration between different types of agents in operation and those used in CPS software (SW) engineering, as well as the effective integration of human goals, remains an open challenge in CPS engineering and operation.

This project aims to create an integrated ecosystem in which partially autonomous AI agents support both CPS development and operation. The research addresses how agent collaboration in the engineering and operation of a CPS can be systematically modeled, observed, and managed, how different agent types and human stakeholders can be integrated, and how effective specification and validation of human intentions can be ensured alongside mechanisms for continuous bidirectional feedback.

To this end, collaboration between agents and lifecycle management, as well as collaboration protocols, will be formalized to enable the integration of AI and non-AI agents into MLOps/SW engineering workflows, embed human goals in Agentic AI processes, and integrate MLOps/SW engineering with CPS operational agents. New methods and Agentic AI-based system architecture concepts will be designed to significantly improve the collaboration of MLOps and SW engineering agents in CPS software engineering. New concepts for proactive SW engineering and MLOps agents will be developed that can observe developers or architects and be seamlessly integrated into IDEs or CI/CD pipelines. Based on this, concepts for the collaboration of SW engineering and MLOps agents with CPS operations agents will then be designed. For both types of agent collaboration, participatory design and LLM-based techniques for capturing and validating human goals will be used. Prototypes will be empirically validated in open-source and industrial CPS environments.

By bridging the gap between CPS engineering and operations through collaborative, partially autonomous agent ecosystems, this project goes far beyond current tool-centric automation and delivers adaptive, more automated, more traceable, and more human-centric CPS software engineering. The novel integration of collaborative Agentic AI architectures and concepts, MLOps methods, software architecture concepts, and human-in-the-loop methods will significantly improve the autonomy and flexibility of next-generation CPS.