Rooms, Not Windows

People are very good at remembering where things are. You remember which room your keys are in long after you forget a file name. Interface research has used this for decades.

Space is how people remember

Malone interviewed office workers about their desks and found that piles are not just storage: a very important function of desk organization is to remind the user of things to do4. Data Mountain showed this works on a screen too, with statistically reliable advantages over a browser’s favourites list2. After an interruption, people use spatial memory to get back to the general area where they were interrupted7. And juggling many tasks at once is hard6.

Workspaces as places

Rooms named the problem window thrashing: spending effort just to keep the right windows visible1. The Task Gallery ran ordinary Windows apps inside a 3D gallery, and its users found it enjoyable to use3. Calm technology added the hallway view: a “Dangling String” that twitched with network traffic and could be read without being obtrusive5.

Agents in Minecraft

Minecraft is already home to many AI agents. Voyager explores and learns skills on its own8. Mindcraft found that talking to each other is the main bottleneck for agent teams10. Generative Agents put twenty-five simulated people in a game town users can watch9.

Omo Space is a mod for Minecraft Java 26.2. You build an agent workspace from ordinary-looking game blocks1112.

How it works underneath

Every workspace is a tmux session (a terminal that keeps running in the background) on your own PC. A small bridge, listening only on your machine, draws those sessions onto in-game screens and sends your keystrokes back11.

Three design rules

1. Leaving changes nothingClosing a screen, walking away or quitting Minecraft never stops a process12.
2. State is reported honestlyIf the session server stops, the screen says INTERRUPTED. Nothing re-runs by itself: “Revive shell” and “Resume Hermes” are buttons you press11.
3. Idle agents sleepAn agent tab that is idle and unseen for 20 minutes exits cleanly and wakes when a screen shows it again11.

Some agents have bodies

A separate project adds a Minecraft player named OMO to the same world. You speak out loud, a voice model hands the job to an agent “brain”, and the brain moves the body through 22 tools13. Two robots, a Microduck and an RP1 humanoid, are physics-simulated in the world and come when called by voice.

A ladder of embodiment. This paper focuses on the first rungs, where the evidence is strongest.

From 25 to 29 October 2026 the builder used Omo Space as his main workspace while streaming live. He ran a real content pipeline from inside the game: ideas, references, recording requirements, editing, publishing and analytics, each in its own room14.

The recordings had already been transcribed by an in-house clipping pipeline (Whisper, small model). Each chunk was read in full by a separate language-model reader, which pulled out claims, quotes, live demos, failures and guest reactions. Claims were then grouped by hand and counted across the eight chunks. A script confirmed that every extracted quote appears in the transcript within three minutes of its timestamp.

A second user, new to Hermes and on Windows, set up his own world from the repository and joined the stream by voice. His account is self-reported.

Seven themes came up in at least seven of the eight chunks. Four of them describe what space does for agent work. We call them moves.

Quotes are as transcribed; the speech recogniser sometimes garbles names.

Move 1 · 8 of 8 chunks

Find: walk to it

The most repeated claim: people can’t remember tabs, but they can remember rooms.

“Nobody remembers which of the 12 terminal tabs are doing what but you can remember space and rooms … It’s the oldest memory trick out there.”

This is Data Mountain’s finding2 applied to agents. At peak the builder ran 21 to 25 sessions at once and still described them as manageable.

Move 3 · 7 of 8 chunks

Monitor: look down the hallway

“Every light that you see in there is an agent still working.”

This is calm technology for agents5. The second user said, without being asked, that watching agents in the world was more fun than watching an AI type in a terminal.

Move 4 · 7 of 8 chunks

Orchestrate: walk the pipeline

The builder’s content business was laid out as a row of rooms: one stage, one agent profile per room.

“We just go to that room and just start building. And when we’re done, then we just leave that room.”

By 29 October he proposed giving every item on his to-do list its own room, calling it a repeatable pattern. The human’s job shrank to reviewing and approving. In the second user’s words: once you press approve inside the game, it just gets posted14.

Two more findings

Every failure below points to a design change.14

1. One agent, one placeGive every long-running agent a stable location and a name you can read from a distance.
2. The place outlives the viewClosing, leaving or crashing the viewer must never stop the work.
3. Show state at room scaleWorking, idle and stuck should be visible from the doorway, without reading text.
4. Be honest about stateNever claim work is running when it isn’t, and never re-run anything on the user’s behalf.
5. Space is not enough for urgencyAmbient signals need a backup that follows you, such as a sound, a toast or an item in your hand, for questions that can’t wait.

Why Minecraft?

Partly culture: it is already multiplayer, already buildable and familiar to a generation of players. Partly technology: the world itself can be written by agents, which earlier 3D desktops like the Task Gallery could not do3. The cost is a game client that is heavier and less reliable than a terminal.

Each move becomes a hypothesis, tested against an ordinary tab setup (tmux or IDE tabs) with 12 and 24 agents.

Some data already exists: bridge logs of sleep, resume and INTERRUPTED events, per-profile agent session stores, and the stream recordings themselves (walking paths, time spent in each room).

Sources

1. Henderson & Card (1986). Rooms: the use of multiple virtual workspaces to reduce space contention in a window-based graphical user interface. ACM TOG. dl.acm.org/doi/10.1145/24054.24056“This results in ‘window thrashing,’ in which the user must expend considerable effort to keep desired windows visible.”
2. Robertson, Czerwinski, Larson et al. (1998). Data Mountain: Using Spatial Memory for Document Management. UIST ’98. microsoft.com/research“Our study demonstrates that the Data Mountain has statistically reliable advantages over the Microsoft Internet Explorer Favorites mechanism…”
3. Robertson et al. (2000). The Task Gallery: a 3D window manager. CHI ’00. dl.acm.org/doi/10.1145/332040.332482“User studies suggest that the Task Gallery helps with task management, is enjoyable to use, and that the 3D metaphor evokes spatial memory and cognition.”
4. Malone (1983). How Do People Organize Their Desks? Implications for the Design of Office Information Systems. ACM TOIS. dl.acm.org/doi/10.1145/357423.357430“A very important function of desk organization is to remind the user of things to do…”
5. Weiser & Brown (1995). Designing Calm Technology. Xerox PARC. calmtech.com“…visible and audible from many offices without being obtrusive.”
6. Czerwinski, Horvitz & Wilhite (2004). A Diary Study of Task Switching and Interruptions. CHI ’04. microsoft.com/research“…the difficulty of shifting among numerous tasks.”
7. Ratwani & Trafton. Now, where was I? Examining the Perceptual Processes while Resuming an Interrupted Task. gregtrafton.com“…participants were able to use spatial memory to return to the general area where they were interrupted.”
8. Wang et al. (2023). Voyager: An Open-Ended Embodied Agent with Large Language Models. arXiv:2305.16291. arxiv.org/abs/2305.16291“…the first LLM-powered embodied lifelong learning agent in Minecraft…”
9. Park et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. arXiv:2304.03442. arxiv.org/abs/2304.03442“…we manifest a society of twenty-five agents in a game environment. End users can observe and interact with these agents.”
10. White, Nottingham, Maniar et al. (2025). Collaborating Action by Action: A Multi-agent LLM Framework for Embodied Reasoning. arXiv:2504.17950. arxiv.org/abs/2504.17950“…the primary bottleneck in collaborating effectively for current state-of-the-art agents is efficient natural language communication…”
11. Omo Space documentation, Agents and terminals. github.com/omoresearch/omo-space “Nothing is ever re-run without you.”
12. Omo Space CHANGELOG. github.com/omoresearch/omo-space “Closing a screen, leaving the chunk, quitting the game or restarting the bridge never stops a process.”
13. Omo, brought to life, README. github.com/omoresearch/omo-brought-to-life “omo-body MCP: 22 tools”
14. Stream analysis: synthesis of seven streams, 25–29 October 2026 (this study). Per-chunk quotes with timestamps are kept in our study notes. “All 791 extracted quotes were re-checked…”

Illustrations were generated for Omo Research (requested model gpt-image-2.5). They are drawings, not screenshots.
The animated walk-through is a diagram of the idea, not a recording.
Omo Research · working paper, draft · 30 October 2026