Rule-Based Software Agent as the Author of a Coordination Issue: Automatic Generation of BCF Files from an IFC Model

Research on Artificial Intelligence (AI) agents in Building Information Modelling (BIM) has recently shifted from the passive querying of models towards systems that invoke tools and act on data. The output of such systems, however, is typically returned as free text, a data listing or a direct modification of the model rather than in the form used by the actual multidisciplinary coordination process, namely the BIM Collaboration Format (BCF). This article asks whether a software agent can generate, without human intervention, a complete and valid BCF file directly from a model in the Industry Foundation Classes (IFC) format, and which decisions this step requires beyond writing the file itself. A synthetic test model was built in the IFC4 schema containing five deliberately introduced defects together with one situation designed as a trap, that is, a compliant element that a context-blind rule would be expected to report. A prototype rule-based agent was developed in Python using the IfcOpen-Shell library. Its detection logic is deterministic and does not use a language model, whereas the whole cycle from reading the model to recording the issues is performed autonomously. The agent produced six issues recorded as a BCF 2.1 file, including automatically derived viewpoints, component selection by global identifiers and view snapshots. The validity of the result was confirmed with an independent BCF reading library and by loading the file into the BIMcollab BCF Manager plug-in running in Autodesk Revit. Five issues corresponded to the actual defects, while one was a false-positive indication resulting from the absence of semantic context in the rule, giving a precision of 0.83 at full recall; since only six issues were generated, this value is illustrative. The results provide a proof of feasibility under controlled conditions: writing a valid BCF file was not the limiting step, whereas the determination of viewpoints, the selection of issue components, the assessment of issue relevance and the recording of machine authorship provenance emerged as open problems.