The Delegation puts multi-agent teams in a 3D office you can walk around

BYOK is not optional, the hosted experience runs on your key

The first thing the README puts in a callout box is that this experience requires BYOK, bring your own key, and that a Gemini API key is needed to run the simulation. Everything else follows from that. The hosted version is linked as a launch link to arturitu.github.io/the-delegation and is described as the full experience, while the repository can be cloned or forked to run locally. Deep integration is what the key buys: native support for text plus multimodal generation through Nano Banana for images, Lyria 3 for music and Veo 3.1 for video. Three consequences follow for anyone evaluating it. First, the demo is not free to run, since every agent turn and every generated asset draws on a key you hold. Second, the key is entered in a page served from GitHub Pages, so it is worth understanding what you are willing to paste into someone else's static site before you do. Third, multimodal output is a paid path, not a preview, which is why the cost panel exists.

Pull request review is the actual product

Version 0.2.0 is built around a workflow borrowed from code review rather than a chat box. Six predefined teams ship as templates, including a Creative Agency, a Film Studio and a PR Agency, and a team editor built on React Flow lets you wire your own multi-agent patterns on a node canvas. Agents work, propose, and wait: the guardrails list names an Auto-approve output option, which means human approval is the default state rather than an add-on, and agents carrying human-in-the-loop properties require your approval before proceeding. Proposals arrive as modals where you can review, comment and merge. Around that sit the visibility features, a team flow visualiser showing the agent hierarchy and task flow in real time, an agent inspector that opens any agent's thoughts, mission and history, kanban and action logs for tool-level interaction, and technical logs exposing raw LLM traces, tool calls and structured agent responses. For educators, that is the payload: the review step is where multi-agent failure is visible.

Per-agent model choice is where the bill arrives

The model assignment is per role, not global. Any agent can be pointed at a different Gemini model, with Flash suggested for speed and Pro for reasoning, so a cheap planner and an expensive reviewer is a configuration rather than a fork. Cost and token tracking is presented as a real-time estimation of usage costs and token consumption for transparency. Two caveats belong next to that claim. Estimation is not a bill, and with three billing modalities in play, text, image through Nano Banana, music through Lyria 3 and video through Veo 3.1, a per-token estimate says little about a run that generated a clip. Gemini is named as the core LLM provider, and there is no second provider in the dependency list, so the per-agent flexibility is flexibility within one vendor's model line. The project also states that it follows official prompting guides for images, video and music, which is a reasonable baseline but also a fixed one.

The office is WebGPU rendering with NavMesh pathfinding

The simulation layer is what separates this from a diagram tool. It is built with Three.js WebGPU under a hybrid GPU/CPU architecture, with autonomous LLM-powered characters sharing one physical workspace. Navigation uses a NavMesh through the three-pathfinding library, and characters find and then claim points of interest such as desks, seats and computers based on whatever task they currently hold, so two agents will not plan to sit in the same chair because the second one has to go looking. A dynamic state machine moves characters between walking, sitting, working and talking, with speech bubbles and facial expressions synced to that state. The assets are custom models and animations rigged in Blender and drawn through an instanced animation system, which is how a crowd stays affordable on modest hardware. State is held in Zustand as a single reactive store shared by the 3D world and the React interface, so the overlay menus projected out of the scene and the panels in the DOM read from the same data.

The dev server listens on every interface

Local setup is two commands, and then a subpath:

npm install

npm run dev

The README points you at the local URL from the terminal, usually http://localhost:3000/the-delegation, and the trailing segment is not incidental: the project is published as a GitHub Pages site under a repository subpath, so the base path is baked into the build and the local run has to match it. The dev script itself is the detail worth reading before you start it:

"dev": "vite --port=3000 --host=0.0.0.0",

That host flag binds every network interface, so the development server answers on your local network rather than only on loopback, with no authentication in front of it. On a shared or office network that is an open door to a development app holding a live API key. The other scripts are unremarkable: build runs vite build, preview serves the production output, clean removes dist with rm -rf dist, which will not work as written on Windows, and lint is tsc --noEmit, a type check rather than a linter. There is no test script at all, so npm test has nothing to run.

MIT for the code, CC BY-NC 4.0 for the office

Licensing is split in two, and the split is stated explicitly. Source code is MIT: all logic, shaders and UI code are free to use, modify and distribute. The 3D models and assets are CC BY-NC 4.0, copyright Arturo Paracuellos, and those are free for personal and educational use but cannot be used for commercial purposes without permission. Since the entire premise is a 3D office full of characters, a team that wants to put this in front of paying users has to license the art separately from the code, and package.json carries only the MIT field, so tooling that reads the manifest will not see the restriction. The package is also marked private, which means there is no npm distribution path to lean on and no registry copy to pull. Neither the licence metadata for the repository nor a single line in package.json captures the non-commercial clause on the assets, and that is the clause that matters in practice.

v0.2.0 shipped on 4 April and the branch went still

Two releases exist: v0.1.0 on 10 March 2026 and v0.2.0 on 4 April 2026, and the last push landed on 16 April 2026. Nothing has moved in the five and a half months since, and the version in package.json is still 0.2.0, so there is no nightly channel and no patch stream to track. The roadmap is a list of unchecked boxes, and they describe the gaps between a demo and a product: an office and 3D space editor with drag-and-drop layout and customisable points of interest, dynamic environment where agents generate props, deeper embodied AI linking reasoning to physical actions, richer context-aware animation, direct human and agent spatial interaction, long-term memory shared across projects, decoupling of core logic from the simulation, and a UI overhaul to unify styling around the project's own brand. The developer note adds context rather than reassurance: version 0.2.0 was built with Google Antigravity as the primary IDE and powered by Gemini 3 Flash, a single-session origin for a codebase with no tests.

Editorial conclusion

The Delegation earns a place in a teaching or prototyping toolkit: it makes agent hand-offs, approval gates and per-role model choices visible in a way a log file does not. Leave it out of production work, since the sample is labelled version 0.2.0, the package is marked private, the 3D assets are non-commercial, and the roadmap items a real deployment would need are still unchecked. Before you build on it, confirm your key's exposure surface, keep the dev server off shared networks, and read the asset licence separately from the code licence.

Frequently asked questions

Official sources

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