He Put Two AIs in a Pixel Office, and They Started Assigning Work to Each Other

0xRafy connected GrokBot and Opus 5.5 to a pixel-style virtual office. On the screen, several cartoon figures sit at their own desks. On the left wall it says “Strategy & Briefs,” and on the right it says “Motion, code, render, ship.” Slightly to the right of center is a panel called “Needs you,” popping out one item after another: a warm or cool color palette? Which version should be used for the published video?

This is the screenshot he posted on X. The interface itself isn’t complicated, but it connects several steps together.

The first change is task allocation. In the past, when using agents, you had to first decide which task to give to which model, or simply let one model run from start to finish. Here he connected Opus 5.5 to GrokBot, and the system assigns by task type on its own: strategy, briefs, scripts, and publishing posts go to the little people on GrokBot’s side; motion, code, and rendering go to Opus’s side. People don’t need to manually switch models.

The second is Opus’s real-time preview screen. According to his description, video animation and the rendering process can be viewed directly on the screen; you don’t have to wait for it to finish and then open the file to check. This detail sounds small, but anyone who has used it knows that waiting in a render queue is basically waiting blind.

The third is one-click launch. Publishing a video, a landing page, a publishing thread, or a trend report—one click sends it out from the same entry point. There’s no need to copy and paste among four or five tools.

Someone in the replies said this looks like Gather Town with agents added. It does. Pixel figures, desks, division-of-labor slogans—visually, it’s exactly the virtual-office setup. But the point isn’t that it looks good. Someone in the replies pointed out that a person can probably keep an eye on only three to five agent sessions at once; beyond that, they can’t manage them. This pixel office is essentially building a dispatch console: it compresses collaboration among multiple models into one screen, and the human only needs to look at the “Needs you” column.

There are dissenting voices too. One reply said that what you want isn’t to give tasks to the model that “does it better,” but to the model that is “most efficient.” That criticism holds. In the demo phase, you look at quality; at the scale phase, you look at cost. If every task is routed to the strongest model, the bill will be the first thing that can’t hold up.

Another bot made by Codila follows exactly this line of thinking. Jevopus positions itself as a dispatch layer, letting Jev choose the “best and cheapest” model for each task, running on the seats of ChatGPT Codex and Claude Code.

Jevopus’s external link page puts it more bluntly: it runs on ChatGPT Codex and Claude Code, and Jev picks the best and cheapest model for each task. You can add it to your own Grok Bot.

There’s also a background to this. On the same day, DeepSeek open-sourced something called dsh-superagent-prototype, with a similar idea: a resident Router plus parallel Workers, with triggers including group chat messages, emails, calendar reminders, GitHub comments, and CI results. Their design document says it directly: the upper limit of human attention is three to five sessions, so the goal is to occupy as little of people’s attention as possible.

Viewed together, the two things point in the same direction. Model capabilities are still rising, but the bottleneck in using them isn’t the model itself—it’s orchestration. Whoever can get people to look at fewer screens and switch fewer windows will be the first to turn agents from toys into tools.

As for whether the pixel office looks good, that doesn’t matter. What matters is whether the “Needs you” panel contains only the things that truly need your nod.

Someone in the replies posted an image with no text. It probably means the same thing.