What happens when you give 12 AI coding agents the task of reconstructing one of San Francisco's most iconic buildings in 3D?
Gemini 4 Carbon just demonstrated something remarkable.
Justine Moore (@venturetwins) asked Google's Gemini 4 Carbon to create a 3D model of the historic Ferry Building using the new /swarm feature.
Instead of working as a single coding agent, Gemini coordinated 12 parallel agents.
They researched the building, planned the reconstruction, wrote Blender Python scripts, and reviewed each other's work through a shared message board.
THE RESULT?
A detailed 3D reconstruction featuring:
• 12,504 individual meshes.
• Approximately 4.62 million triangles.
• The Ferry Building's main hall and glass-roofed interior.
• Market stalls and architectural details.
• The waterfront and surrounding piers.
• A walkthrough of the reconstructed environment.
• Material and clay-render views revealing the underlying geometry.
BUT THE MOST INTERESTING PART IS THE WORKFLOW.
Gemini 4 Carbon didn't simply generate a beautiful image of a building.
It coordinated an entire team of AI agents working on a shared technical project.
Here's how the experiment worked:
• Parallel research: agents gathered architectural references and information about the building.
• Task decomposition: the swarm divided the reconstruction into smaller assignments.
• Blender scripting: agents wrote Python code to construct the scene programmatically.
• Shared coordination: a message board allowed agents to exchange findings and review each other's work.
• Agent-to-agent criticism: different agents evaluated the output and challenged decisions made by others.
• Integrated 3D scene: the resulting geometry was assembled into a detailed environment.
WHY DOES THIS MATTER?
Complex 3D production involves much more than modeling individual objects.
It requires research, architecture, scene organization, geometry, materials, integration, and quality control.
A single AI agent can struggle to maintain all that context while working on a large project.
Multi-agent systems offer another approach.
Different agents can specialize in separate tasks, work in parallel, and review each other's results.
This is starting to look less like AI generating isolated assets and more like autonomous software and 3D production teams.
Of course, 4.62 million triangles don't automatically mean the model is historically accurate, optimized, or production-ready.
The demonstration also doesn't establish the total computing cost or how much human supervision was required.
But the ability to coordinate 12 agents on a complex Blender reconstruction is impressive in its own right.
And it raises an interesting question.
If AI agents can collaborate on an entire building today, how far are we from having them construct complete virtual cities?
ORIGINAL EXPERIMENT
Justine Moore (@venturetwins):
https://x.com/venturetwins/status/2122043781507948954
Google Antigravity:
https://antigravity.google/
Blender:
https://www.blender.org/
Would you rather use one powerful AI coding agent or a swarm of specialized agents for complex 3D projects?
#Gemini4Carbon #Blender3D #AIAgents
YouTube / ABVX
