A Glass Tower for AI Agents: See at a Glance Who's Stuck and Who's Waiting on You

What's it like to run a dozen AI agents at once?

A pile of chat windows in the browser. To find out who's waiting for your approval, who's stuck, or who's burning through budget, you have to click into each one. 0xRafy offered a solution on X: GrokBot Tower, a prompt-generated agent management console that looks like an isometric glass office building.

You can tell who's stuck from outside the building

Each bot has a status ring under its feet. Green means working, amber means waiting on you, red means stuck (it flashes and smokes), and gray means resting.

The floors also have LED strips. If someone on a floor is waiting on you, the strip turns amber; if someone is stuck, it flashes red. One glance at the building tells you which floor has trouble today.

The elevators move, carrying bots up and down. The repair bot on the Ops floor takes the elevator to fix a stuck agent, throws sparks while fixing it for about 3 seconds, then returns.

Six floors, each with a different vibe

From bottom to top: Ops, Research, Content, Studio, Code, HQ.

HQ has a main desk and a KPI wall screen; Code has server racks and neon; Studio has a render preview screen and beanbag chairs; Content has a podcast booth with an ON AIR sign on the door; Research has bookshelves and a live map; Ops has an entire wall of monitors.

Click a floor and the floors above lift and fade out while the camera smoothly pushes in to that floor. Scroll to zoom, drag to pan, double-click to reset.

Three panels to manage everything

The top bar has a live clock, agent count, and five KPI pills (working/waiting on you/stuck/resting/today's spend); the pills themselves work as filters. The search box dims agents that don't match.

The left column is the to-do list. Stuck items go first, explaining clearly 'why it's stuck,' such as 'X API rate limit (429),' followed by a fix suggestion: 'Switch to the backup key.' Questions have Approve, Reply, and Skip. When you click Approve, the bot in the scene jumps and a heart pops up.

The right column is the context card. Click the building to see the whole-building grid, click a floor to see the team, and click a person to see task steps and cost. The spend panel has a budget bar and today's forecast, highlighting the agent burning the most money right now.

Chat can take action directly

The main assistant's 'Ask' has several quick chips: Summary, Who's waiting, What failed, and How much was spent today. You can also type your own, such as 'What is Atlas doing?'

Answers are computed from the live simulation and come with action buttons. Press Approve and that row turns green, collapses, and the count updates. When everything is handled, it shows 'All handled.'

Each agent also has its own chat, where you can answer its questions or call for repair.

How to build it

Paste the author's prompt into GrokBot. Write your agents in a block at the top: name, floor, job, model. Leave the rest to it.

It generates a self-contained HTML file with glass floors, elevators, panels, and bots. It tests approval, repair, chat, search, mobile layout, and dark mode on its own, and fixes any issues it finds. In the end, it gives you a link.

The author says 15 minutes.

The honest truth in the comments

Someone replied: At this point, what we're missing isn't more AI agents; it's a manager to manage the managers.

Another added: Yesterday it was an office, today it's a tower, and by Friday it'll be an entire robot city—complete with traffic jams.

Those two lines are more interesting than the feature list. Once you have many agents, the person doing the scheduling becomes the bottleneck. Putting agents into a building may not solve the management problem, but at least you don't have to keep switching between a dozen windows. Who's stuck and who's waiting on you—the building will tell you.

In another reply, codila shared Jevopus, which they built: it packages Claude and GPT into one agent so Jev can pick the cheapest suitable model to run tasks. That's also agents managing agents.