Name of participant: Philine Kowol
Project’s name: OrKA: Orchestrator for conversational agents in multi-agent systems
Project description:
The OrKA project is researching and developing a prototype orchestrator agent for text-based multi-agent systems. The aim is to enable specialised conversational agents to collaborate within large, heterogeneous system environments whilst ensuring a consistent user experience.
A growing number of specialised AI and dialogue agents are being developed for various products, brands, services and internal processes at companies such as Bosch. These agents are often developed independently of one another and are based on different technologies, frameworks or communication interfaces. This results in fragmented system landscapes characterised by redundant functions, inconsistent user experiences and maintenance issues.
The project therefore examines how a central orchestrator agent can intelligently distribute tasks to suitable agents, coordinate multiple agents, and consolidate responses into a consistent interaction. In doing so, it conducts a comparative analysis of various orchestration approaches, including rule- and workflow-based methods, graph-based approaches, and LLM-based strategies. Another focus is on integrating new agents as easily as possible through standardised self-descriptions, automatic validation steps and interoperable communication interfaces.
In addition to prototype development, the following scientific questions are being investigated:
- How much autonomy should individual agents have?
- How visible should the multi-agent structure be to users?
- Which tasks are better suited to LLM-based decision-making processes, and which are better suited to deterministic methods?
- How can heterogeneous agents be integrated in a robust and scalable manner?
- Which orchestration strategies are most suitable?
- How can conversational multi-agent systems be systematically evaluated?
The evaluation is based on realistic usage scenarios from the industrial sector. Among other things, it analyses quality, robustness, latency, traceability and user experience. The project combines cutting-edge research into multi-agent systems, conversational AI and LLM orchestration with practical requirements from industry.
Software Campus Partner: TU Berlin & Bosch
Implementation period: 01.01.2026 – 31.12.2027




































![[KOM,BI]Co-citation-based machine learning to determine promising research projects](https://softwarecampus.de/wp-content/uploads/2022/02/KOMBI-768x768.png)




































































