Agentic LLMs for Conceptual Systems Engineering and Design
Research output: · Journal of Mechanical Design
RolePostdoctoral researcher and first author
Background
Early engineering design is iterative and stateful. Requirements change, components interact, models must execute, and decisions need to remain traceable. A fluent generated answer is therefore not enough to demonstrate a useful engineering workflow.
Method
This work introduces a JSON-serializable Design-State Graph that stores requirements, physical embodiments, and Python physics models as connected design nodes. It compares a nine-role multi-agent system with a simpler generator–reflector baseline.
The study was developed with Mark Fuge at the IDEAL Lab, ETH Zurich.
Results
- 60 runs across two model families, two agent architectures, three temperatures, and five random seeds.
- Both architectures maintained valid JSON and embodiment tagging in the reported experiments.
- The multi-agent system produced more granular design-state graphs.
- Requirement coverage remained below 20%, and average code compatibility remained inconsistent, especially for the larger multi-agent workflow.
Limitations
Structured orchestration can improve decomposition and retained state, but more agents do not automatically produce better engineering. Requirements fidelity, executable modeling, and reliable completion remain open research problems. The released traces make those limitations inspectable.