A recent arXiv paper proposes treating LLM coding agents as suppliers of physical system models. In this setup, a client hands the agent a description of a device, sometimes accompanied by recorded data or a target response, and the agent must produce a model that meets the handover requirements.
The authors frame the task as a procurement process: the model is accepted only if it satisfies the specified criteria, and calibration is used to align the model with the given data or target. The work is demonstrated in Modelica, a language commonly used for modeling physical systems.
This approach could make model generation more automated and reliable, though the abstract does not detail the specific results or limitations. The key contribution is the formal framing of LLM agents as accountable suppliers, rather than just code generators.