I read five papers on agentic AI in the built environment this month and wrote each of them up separately. Reading them together is more useful than reading any one of them, because the architectures have almost nothing in common and the conclusions keep converging.
Five different answers to the same question
- An LLM orchestrating Revit and Blender through MCP servers, closing the loop from site photo to updated 4D model.
- A multi-agent system with a publish-subscribe bus and shared memory, producing cost distributions instead of point estimates.
- A code-executing agent that explores an IFC model at runtime rather than assuming its structure.
- A fine-tuned 7B model, trained with reinforcement learning, generating regulatory rule skeletons.
- A structured-regulation pipeline that reads building codes directly and decides what applies.
MCP orchestration, multi-agent coordination, iterative code execution, small-model fine-tuning, structured pipelines. Five architectures. None of them is a bigger model.
The pattern
Every one of these wins through orchestration, tool access and a feedback loop. The two results that state it most directly are worth putting next to each other.
In the TUM study, the weaker model using the better paradigm beat the stronger model using the worse one. In the cost estimation study, handing the whole problem to one strong model in a single prompt performed worse than a deterministic spreadsheet method.
Two independent teams, two unrelated tasks, same finding: how you arrange the work beats which model does it.
Everyone measured the same thing first
Every paper that compared against a baseline led with hours saved and manual interventions avoided. Accuracy came second.
That is not because accuracy matters less. It is because time is the number a client can act on immediately, and a 12.5% error rate means nothing to someone who has never been told what their current error rate is. If you are writing a proposal, lead where the research leads.
Data quality is the ceiling, twice over
Two of the five hit the same wall from opposite directions.
In the TUM work, models that had been checked and corrected in Solibri scored noticeably higher. The agent was not the limit; the model was. In the cost estimation study, variation in the source data caused 58% of the errors.
That makes getting models into a defined, checked state a legitimate service in its own right rather than housekeeping you do before the interesting work. If a client's data sets a hard ceiling on what any agent can achieve, raising the ceiling is the first engagement, not a precondition you complain about.
Nobody claimed autonomy
The best result in the set is 55 to 57% strict accuracy on a benchmark. The most impressive demo is one wall built from miniature bricks. The strongest compliance numbers measure interpretation rather than end-to-end correctness.
Not one of these papers claims deployment-ready autonomy, and the honest pitch that follows is decision support. An agent that does the first pass and routes what it is unsure about to a person. That is also where liability sits comfortably, which matters more in this industry than in most.
What I would tell a firm this week
- Do not buy on model brand. The research says the arrangement of the work matters more, and that is a question about the vendor's engineering, not their API contract.
- Ask what it was measured against. Three of these five compared against a real baseline. Most products compare against nothing.
- Fix the models first. It is the cheapest available uplift and it improves everything downstream at once.
- Scope for supervision. Every credible result in this literature has a person in the loop. Any proposal that does not is either overselling or has not been measured.