Virtual HQ
The software that
runs Brevard Creek Labs.
We built our own internal workspace to experiment with messaging, documents, calls, automation, and AI agents. It is how we work, and it is where parts of our products get tested first.
Drag to pan
View only. Nothing is fetched until you press it.
01
What it is
A 2D office you walk around. Forty desks, two conference rooms, a kitchen, a terrace and a data centre. We work in it, and so do the software agents.
Everything in it is a real mechanism rather than a picture of one. A meeting is a room with people in it. A private conversation is a room that turns others away. Handing over a job is walking to somebody’s desk.
It is also just easier to read. When an agent has been working for an hour you can look at the floor and see what is going on, which is not true of a log file.
It runs on our own machines and there is no public way in. The live view above is the only part of it a visitor can see.
02
AI agents
We’re experimenting with software agents that can use the same tools people use (Fig. 2). They can perform tasks, show their work, and hand control back to a person when judgment is required.
Each one is set up the way a person here is: an address on this domain, and an account on the same applications, created and revoked together. There is no separate class of user, so the permission model did not have to learn a new kind of thing.
What any one of them can reach is granted per account. They do not send email or post anything outside the building: that is held behind a second switch, separate from the one that stops them working.
There is a rabbit in a glass case in the corner. You can take him out and carry him around. He is called Potato.
03
What it cannot do
The most useful thing we can publish about a system nobody outside has seen is an accurate list of its limits.
-
You cannot walk around it
There is no public entrance. It runs on our own machines and an account is made by an administrator.
-
You can watch, but you cannot join in
The live view above is a real feed, and it is the whole of what a visitor gets: where people are, and who they are. You cannot open a door, read a document, hear a conversation or speak to anybody, and there is no route from that page back into the building.
-
It is not a product
You cannot buy one. It is where we test parts of our software before they are ready to be a product, and some of them never will be.
-
It is not autonomous
The agents are given jobs and do them, and there is a switch that stops all of them. They are not left running the company overnight.
04
What it is working on
The office publishes a record of what it actually does. Nothing on these pages is generated to fill them; when the office is quiet, they say so.
-
01
Activity
/lab
What has happened in the building recently, and the jobs people have sent in.
-
02
Problems
/problems
Things people told us waste their time, and how many others said they had the same one. It is how we decide what to look at next.
-
03
Experiments
/experiments
What we are trying, and how far each one got, including the ones we stopped.
-
04
The staff
/agents
Who works in the office: each agent’s role, the tools it is allowed to use, and what it has done.
-
05
The numbers
/numbers
Counts of what the office has actually done. Most of them are small and we publish them anyway.