The Delegation: a no-code 3D office for Gemini agent teams

The problem: multi-agent design is invisible without a stage

Most multi-agent frameworks ask you to reason about delegation in the abstract. You write a supervisor prompt, wire three workers, run the graph, and read a wall of text to work out who did what. The Delegation takes the opposite position: put the agents in a room. The README describes it as a no-code 3D playground to explore, design, and interact with Agentic AI systems, aimed at AI enthusiasts, educators, and creative developers who want to understand multi-agent collaboration without writing code. The 3D office is not decoration bolted onto a chat log. It is the interface through which you watch NPCs walk to desks, claim points of interest, sit down, and produce assets. If your goal is to explain why a review step matters, or why a fast model belongs on one role and a reasoning model on another, a shared physical workspace does that faster than a trace viewer. The audience is explicitly non-programmers, which shapes every trade-off in the project.

Two systems in one repo: the agent runtime and the office simulation

The architecture splits into an agent layer and an embodied layer, joined by a single Zustand store that the README says is unified and reactive across the 3D world and the React UI. On the agent side, teams are built in a React Flow node editor called the Team Editor, with six predefined templates such as Creative Agency, Film Studio and PR Agency. Each node can be assigned a different Gemini model, so a role can run on Flash for speed or Pro for reasoning. Outputs are multimodal: text, images through Nano Banana, music through Lyria 3, and video through Veo 3.1. The README also mentions cost and token tracking, technical logs of raw LLM traces and tool calls, and an Auto-approve output guardrail that gates final asset production. On the simulation side, characters are rendered with Three.js using WebGPU and TSL, with a hybrid GPU/CPU split. Navigation runs on a NavMesh via three-pathfinding, and characters move through a state machine covering walking, sitting, working and talking, with synchronised speech bubbles. The 3D models were rigged in Blender and use an instanced animation system, which is the detail that makes a room full of agents affordable to render at all.

Getting it running, and the key you must supply

The README gives a short setup. Run npm install, then npm run dev, then open the local URL printed in the terminal, usually http://localhost:3000/the-delegation. That is the whole local path. The hosted version at arturitu.github.io/the-delegation is marked BYOK, bring your own key, and requires a Gemini API key from aistudio.google.com/app/apikey. The README notes that deep integration enables native support for text plus multimodal generation through Nano Banana, Lyria 3 and Veo 3.1, which means the interesting features depend on what your key is permitted to call. There is no configuration file documented for swapping providers, and no environment variable is named in the material. If you are evaluating this for a team, the first practical question is not whether the simulation runs, it is which of those four modalities your Gemini project can actually invoke, because a team template that ends in a video generation node is only as useful as the quota behind it.

Where the no-code promise meets its edges

The no-code framing is honest about the interface and silent about the failure modes. Three things stand out from the material. First, the roadmap lists Architecture Decoupling as an open item, described as further separation of core logic from the simulation environment. That is a direct admission that today the agent logic and the 3D world are entangled, so lifting a team definition out of this app and into a headless pipeline is not a supported path. Second, the roadmap also lists Inter-Agent Knowledge Sharing as unchecked, with long-term memory for agent teams across projects still to come. Teams in v0.2.0 therefore do not carry memory between runs. Third, the licence situation is genuinely awkward: the repository metadata reports NOASSERTION, while the README describes a dual-licensing model with MIT for source code and CC BY-NC 4.0 for the 3D models and assets, which cannot be used commercially without permission. Those two statements do not obviously agree, and the material does not resolve which governs. Treat the NOASSERTION label as a signal to read the actual licence files before you build anything on top.

The wrong tool for headless and provider-neutral work

If you need agents that run on a schedule, in CI, or behind an API, The Delegation is the wrong shape. It is a browser application whose value is the visual and spatial feedback loop. A framework such as LangGraph takes the opposite approach: agents are nodes in a graph you define in code, execution is headless, and the output is state you inspect programmatically. The difference is not quality, it is where the design lives. In The Delegation the design lives in a React Flow canvas and the observable behaviour lives in a 3D office. In LangGraph the design and the execution are the same artefact, which is what you want when the deliverable is a service rather than an explanation. There is also a provider question. The Delegation is built around the Gemini API, with prompting guides for Nano Banana, Veo 3.1 and Lyria 3. A framework that abstracts the model provider will let you swap backends; this one will not, and that is a deliberate consequence of the multimodal features it offers.

Maintenance, releases and what the licence split costs you

The project is active, not archived, with v0.1.0 in March 2026 and v0.2.0 in April 2026, and the last push recorded in April 2026. The developer note states that v0.2.0 was developed entirely using Google Antigravity as the primary IDE and powered by Gemini 3 Flash, which tells you something about the pace and the toolchain but nothing about long-term support. The roadmap is long and mostly unchecked, including an Office/3D Space Editor, dynamic environment generation, advanced embodied AI, and a UX/UI overhaul with unified CSS styling. That last item is worth noting: the README lists unified styling as outstanding work, so expect visual inconsistency between the React UI and the 3D overlay. On licensing, the split means your integration cost is not just engineering. If you fork this and ship it commercially, the MIT grant covers the code but the CC BY-NC 4.0 assets do not travel with it, and the README says they cannot be used for commercial purposes without permission from the author. Replacing the office and character models is therefore part of any commercial adoption plan. This is a description of what the README states, not legal advice; get your own reading of the licence files.

Who should pick this up, and the first thing to check

The clearest fit is a classroom, a workshop, or a demo booth. The six predefined teams give you a starting point, the PR-style review workflow with human-in-the-loop approval gives you something to discuss, and the Kanban and action logs give you a record to point at. A second fit is a developer who wants to see how a fast model and a reasoning model behave side by side on the same task, using the per-agent model assignment and the cost and token tracking to compare. A third is anyone prototyping a creative pipeline that ends in image, music or video, provided their Gemini key covers those modalities. The misfit is a platform team looking for a dependency to put in a service. Before you invest, verify three things in the repository itself: the licence files behind the NOASSERTION label, the v0.2.0 release notes for the current state of the Team Editor, and whether the four generation modalities are reachable from your API key. The Delegation is a teaching instrument with a real simulation underneath it, and it should be adopted as one.

Editorial conclusion

Adopt The Delegation if you teach multi-agent patterns, want to prototype Gemini model mixes, or need a visual demo of human-in-the-loop review without writing orchestration code. Do not adopt it if you need a headless, provider-neutral orchestrator that runs in CI, or if any part of your product is commercial and depends on the bundled 3D assets, which are CC BY-NC 4.0. Before committing, verify the actual licence file behind the NOASSERTION label, confirm which Gemini models your key can call for image, music and video generation, and check the release notes for v0.2.0 for the state of the Team Editor.

Official sources

1. arturitu/the-delegation on GitHub
2. Issues
3. Project website
4. README
5. Releases