Hi all,
I tested Agentic Modeling (Veras for Revit 5.2.2.1, GPT-6 Astra, effort High / Extra High) on the existing Neo-Baroque limestone facade of a 1890 building in Munich. Overall impressive: with coordinates relative to the internal origin, levels and an axis chain in the prompt, walls, corner radius and openings landed within millimetres of our surveyed Revit model. A detailed master bay (cast-iron casings, balcony slab, balustrade, rustication, window surrounds) came out as a useful simplified stand-in. Three points would make it much more usable in a BIM workflow:
1. Baked solids are not editable
Sweeps, extrusions and revolves bake as Generic Models whose geometry cannot be edited. It would help a lot to bake them as In-Place families or loadable families built from native Revit forms (extrusion, sweep, revolve, blend), so the result can serve as an editable template. In one interior test the agent produced real window families, in the facade test only Generic Models.
2. Families are imperial in a metric project
All loadable families arrive in imperial units, and the scene JSON shows “units”:“feet” while the agent reports in metres. Please create families in the project units. Every conversion step is an error source.
3. Bug: source image list is frozen in a running agent session
Removing or adding source images in the UI is not reflected in the running agent session, even after restarting Veras. The agent keeps listing the original 12 images, while the UI shows 8. Adding new images then fails with the “more than 12 images” message. Only a new session helps.
4. Bug: Spell checking in prompt Window
On German Windows istallation English language prompts are unreadable due to the german spell checking marking every word. I posted about this issue before. Please add a language switch for imput fields in settings.
Smaller wishes
- Copy/paste for the agent console (reports and assumption lists).
- Pan in the 3D viewer for checking positions (Revit 3D Navigation)
- Let the agent read host grids and levels as data (not only names/elevations), so registration does not have to be passed as coordinates.
- Note in the docs that source images are downscaled to about 1024 px for the agent, so users prepare crops instead of large overviews.
Happy to share more details or test data.
Christian Meinert, Aukett + Heese, Germany