Complex engineering workflows such as vehicle design can be tackled with multi-agent systems (MAS) that orchestrate specialized tools. Most implementations are message-based, where agents extract necessary information from the messages created by the agents and their tools. While very flexible and "natural" in a large language model (LLM) setting, message-based systems often face challenges in complex engineering workflows: the parameter space can be large, the dependencies of the parameters can be complex and the whole flow can involve many steps, making it difficult for the agents to find and use the correct parameter information or taking the correct next step. Our work motivates using state-based MAS as a design pattern for complex engineering workflows. In a state-based approach, tools and agents work with and directly update the shared system state, eliminating the need to pass all information via messages. Therefore, the messages exchanged by agents can focus on important information and what to do next instead of passing parameters. Furthermore, it enables detailed deterministic control over parameter dependencies, e.g. invalidating parameters on changing inputs, allowing for a better handle on system dynamics besides prompt engineering. Thus, state-based MAS help in achieving a more accurate and reliable system at the expense of some flexibility and the additional effort to explicitly handle the state. We propose a practical approach to implement such state-based MAS: (1) create a model for the variable dependencies using expert knowledge or a sensitivity analysis; (2) derive a minimally coupled dependency graph by eliminating not strictly necessary links; (3) use this graph as the blueprint for the state-driven MAS layout and implement it with the necessary tools to handle the state, like a tool that handles state updates by the user. This results in a robust minimally coupled state-based MAS suitable for complex engineering workflows. REFERENCES [1] Wooldridge, Michael. An introduction to multiagent systems. John wiley & sons, 2009. [2] La Malfa, Emanuele, et al. "Large language models miss the multi-agent mark." arXiv preprint arXiv:2505.21298 (2025). [3] Brafman, Ronen I., and Carmel Domshlak. "From One to Many: Planning for Loosely Coupled Multi-Agent Systems." ICAPS. Vol. 8. 2008.