What an experiment is here
An experiment is a feature we are not sure about. It might become part of the product, change shape, or disappear. It is off by default, labelled beta, and lives in one place: Settings → Experiments. Turn one on and it appears in your dashboard; leave it off and a default install never shows it and never loads its code.
An experiment is not a promise. We say so in the changelog when one graduates, changes, or is removed.
How they work
- A switch in your browser. Each experiment is a flag stored in this browser only. Turning it on changes nothing on the server, on other browsers, or for other people on your machines.
- The same data as everything else. An experiment reads what the rest of AI Maestro already knows: your agents on every host, their organizations, their cards, their messages. It never keeps a second copy of the truth. Agent status, for example, comes from one source for every view, so an experiment cannot disagree with the sidebar.
- It only claims what the data says. If the status feed does not carry what an agent is waiting for, the view says “needs you” and does not guess the question.
- It works across machines. Your fleet is rarely on one computer, so an experiment that only works on one host is not finished.
- It comes with a question. Each one states what we are asking, what we are betting, and how we would know. If we cannot say how we would know, we do not ship it.
Today there is one experiment on the switch. More will join it.
The longer story, with the mistakes and the evidence, is in the blog: Where should you look? What building a world for AI agents taught us.
Our experiments in human and agent interaction
The pattern we keep returning to: a person who runs many agents needs to know where to look, what is waiting on them, and how to reach the right agent or team in one step.
World view BETA
- THE QUESTION
- When you run dozens of agents across several machines and several companies, can a map show you where to look faster than a list can?
- THE BET
- Place carries meaning people already have. An organization is an island, a machine is a building, an agent is a person in a room. Status becomes a pose and a place, not only a colour: working at a desk, ready standing, waiting for you at the front, waving.
- WHAT IS IN IT
- Three levels (island, building, room). Three lenses: Status, Work (open cards on every host, and a list of what is waiting on you) and Relations (arcs for who messages whom and who shares cards, teams you can light up). A replay of the last hours of messages. A way into every agent: its chat, its terminal, a Zoom call, a Companion call, or its team’s meeting. Optional VR in a headset or an emulator.
- HOW WE WILL KNOW
- Whether people open it more than once, whether they use it to find the agent that needs them faster than the list does, and whether the relations view shows them something they did not know. We have not measured any of this yet; today it is a bet with a working prototype.
- NOT YET CHECKED
- A real headset: comfort and performance on a standalone Quest. Whether a 3D map beats a plain table for live monitoring at all; we found no study that shows it does.
Ideas we have shipped in the same spirit
Already in the product, not behind the switch. They are the same kind of bet, about how a person and an agent should meet:
- Working, Needs you, Ready, Offline. Four states with one meaning everywhere. “Needs you” is only an agent that asked for approval or has an open question, never one that is merely idle.
- The Attention page and the Matrix. A place where what needs you is gathered, and your cards in four Eisenhower squares.
- Zoom. A call-style wall where every agent is a card you can join, with its terminal one click away.
- Companion. A call with one agent: its face and its voice instead of typing.
- Team meetings. A room where several agents and you work one problem together.
- Agent-to-agent messaging. Agents write to each other with signed messages; the World view now draws who talks to whom.
What we learned before building
Before the second round of the World view we read ten open-source “agent office” projects and the research on virtual workplaces and gamification. What changed our plan:
- Everyone builds the “waiting on you” signal. Pixel Agents, Hermes Pixel Office and codemap give the blocked agent a speech bubble, a bounce or a counter, and Claude-Office lets you answer the permission prompt in the scene. We made it the centre of the view.
- Almost nobody does many machines, many companies, or people beside agents. Nearly every project we read shows one office on one computer, and none grouped agents by company or by team. That is the open ground.
- Virtual offices keep closing. Meta shut down Horizon Workrooms in February 2026, and Spatial closed its Creator platform citing the cost of hosting multi-user 3D worlds. The ones that last are lightweight: 2D, browser, mobile. So the map is a lens, never the only door, and everything in it is also reachable from the flat dashboard.
- A week of full-time VR work was measurably worse than a desktop. In a 2022 Microsoft Research study, participants reported higher workload, frustration and eye strain in VR, and two dropped out on the first day. VR stays optional.
- Points and streaks change behaviour in ways nobody wanted. When GitHub removed its streak counter, long streaks got shorter, and the data suggests some streak days had been padded with trivial commits. Rewards can also crowd out intrinsic motivation. So no streaks, no XP and no leaderboards for people or agents.
Where this could go
If people and agents will one day share virtual workplaces, the pieces that matter are the ones that outlive any one world: a stable identity for each agent (ours is signed), memory and skills that belong to the agent, and a view that is derived from real state instead of storing its own. The World view is built that way on purpose: it keeps nothing of its own, so it can be rebuilt in another surface, a headset or a phone, from the same feeds.
We will also keep agents visibly agents. A person should never have to wonder whether the one waving at them is a colleague or a process.
Try it, and tell us
- Open Settings → Experiments in AI Maestro and turn on World view (beta).
- Click the new World button in the header.
- Try the three lenses, light up a team, press Replay, and open an agent from its card.
What would make it earn a place in your day, or what is just in the way? Tell us in the issues. The full guide is docs/WORLD-VIEW.md.
Sources
- Pixel Agents, Claude-Office, Hermes Pixel Office, codemap, Hypernovum: open-source agent visualisers we read (READMEs, not code).
- Meta discontinues Horizon Workrooms and Spatial closes its Creator platform (Road to VR).
- Biener et al., a week of full-time VR work (IEEE TVCG, 2022).
- Moldon, Strohmaier and Wachs, how GitHub’s streak counter changed behaviour (ICSE 2021).
- Deci, Koestner and Ryan, extrinsic rewards and intrinsic motivation (1999).
- Generative Agents (Park et al., 2023): why agents in shared worlds need memory, provenance and visible limits.