A 3D office for coding agents is a real coordination bet, not just a gimmick

Agent Office is a 3D multiplayer environment where coding agents occupy physical desks and human collaborators walk around watching them work. Each GitHub repository becomes a floor of a building. You ride an elevator to pick a project, walk up to an empty desk, press E, and assign a Claude Code, Codex, or OpenCode worker to it. That agent's live terminal appears on the laptop model in front of the desk. When the agent needs input or finishes a task, its character jumps up and down and dings — a spatial notification system that replaces the usual wall of status badges.

The spatial metaphor is doing real work here, not just decoration. GitHub issues and pull requests hang on cork boards on the walls. You can hand an issue directly to a worker, give it a dedicated git worktree, and open a PR with a single keypress. Multiple humans share the same space with voice, chat, and a shared whiteboard on the lounge TV. Each person's character carries their name, and their terminal interactions are attributed to their account — so comments, merges, and pushes appear on GitHub under the right person's identity.

Deployment is thought through carefully. The AWS path spins up a t3.xlarge instance with a 50 GiB disk and a fixed Elastic IP in roughly two minutes, but the office never exposes itself to the public internet — everyone reaches it through SSH tunnels, which also means voice and screen sharing work without certificate management. The provision script runs the office under systemd so workers survive a server restart. For teams, each person gets their own account with their own Claude subscription and GitHub credentials stored separately under .agent-office/homes/<account>/.

The author is honest about scope: this is built for one person's workflow and changes fast. Keys move, screens get redrawn, features appear and disappear. The README says to fork it and bend it into what you need — which is the right framing. This is not a polished product for general distribution; it's an opinionated personal tool that happens to be open.

What makes it interesting is the bet it's making: that spatial presence and shared observation of agent terminals changes how a small team coordinates around autonomous coding work. Whether that bet pays off depends entirely on whether the overhead of navigating a 3D space is lower than the overhead of checking status in a dashboard. For solo use the answer is probably no. For a team running several agents simultaneously on different repos, the spatial model — where you can literally see which desk is jumping for attention — might actually compress coordination cost in a way a flat list of statuses doesn't.