Agentic mode: the AI can now see its own designs
Highlights
- Agentic mode is now the default: the AI renders, inspects, and revises before finalizing
- Upgraded the default AI model behind Standard chats
- Benchmarks show a significant quality jump over single-shot generation
- Failed generation attempts no longer overwrite your last working design
- Lower costs for agentic runs
What’s New
The AI can now see its own renders
Until now, the AI generated OpenSCAD code in one shot and hoped for the best. With agentic mode, it gets proper eyes on its own work: after writing code, it compiles the design, renders calibrated views of it, and inspects the resulting mesh for problems like non-manifold edges, open shells, or disconnected bodies. It reads that evidence and revises before showing you anything.
You can watch this happen live in the event log under the viewer: the agent designs a candidate, renders it, validates the mesh, and either finalizes it or tries again.
Significant quality gains
In our evaluation set, single-shot generation averaged 84 out of 100. Giving the AI calibrated multi-view renders and mesh diagnostics to check its own work raised that to 94 out of 100, with fewer missed dimensions and fewer broken meshes. That gap is measured on simpler designs. On more complex ones, where a single-shot attempt has more room to go wrong, the gain is at least twice as large.
A more capable AI model
Standard chats now run on an upgraded AI model, paired with the agentic loop described above, replacing the old single-shot path.
More reliable results
- If a generation attempt fails partway through, the session now ends cleanly instead of leaving a broken half-response
- Your last known-good design is preserved. A failed retry can no longer silently replace a design that already passed validation
Lower costs
- Reduced pricing for agentic runs
- Follow-up messages in the same conversation now cost less to process, thanks to better context reuse
