Your AI agents now have a 3D office and a PR review queue

What it does

The Delegation is a browser-based playground for assembling and running multi-agent AI teams without writing code. You lay out agents and their relationships in a React Flow node editor, assign each a specific Gemini model, and dispatch them to generate text, images, music, or video. The twist is that these agents don’t just return JSON; they inhabit a shared 3D office built with Three.js WebGPU, where they pathfind to desks, sit down, and flash speech bubbles while they work.

The interesting bit

Instead of stopping at a flowchart, the project adds an embodied simulation layer—complete with NavMesh pathfinding, Points of Interest for desks and chairs, and state machines for walking, talking, and working—so you can literally watch your Creative Agency team claim a computer. It also bakes in PR-style human-in-the-loop approvals and per-agent cost tracking, which treats agent orchestration more like a dev workflow than a black-box pipeline.

Key highlights

• Node-based team designer with six predefined industry templates (Creative Agency, Film Studio, etc.).
• Per-agent Gemini model assignment and real-time token/cost estimation.
• Multimodal output generation via Gemini: text, Nano Banana images, Lyria 3 music, and Veo 3.1 video.
• Embodied 3D simulation with NavMesh navigation, dynamic state machines, and synced UI overlays.
• Human-in-the-loop guardrails with simulated PR review modals and auto-approve toggles.

Caveats

• Requires your own Gemini API key to run; there is no offline mode or fallback provider.
• Custom 3D office and character assets are licensed CC BY-NC 4.0, so they cannot be used commercially without permission.
• Several headline features—an office editor, long-term inter-agent memory, and dynamic environment modification—remain unchecked on the v0.2.0 roadmap.

Verdict

Grab this if you are an educator, creative developer, or AI enthusiast who wants to make abstract agent graphs feel concrete. Skip it if you need vendor-agnostic LLM support, fully open commercial assets, or a finished feature set.