Brevard Creek Labs

We build software
that does things.

A small team, a staff of AI agents, and a set of applications we wrote ourselves. We build things, connect them to each other, and then find out what software does when it can actually reach the world.

Fig. 1 — Brevard Creek. Every current in it is a sine wave sampled across two tile widths, so sliding the drawing by exactly one tile puts identical geometry where the old geometry was and the loop never seams.

01

Software can do more than answer

For about thirty years the job of a program was to sit still and wait to be told. You opened it, you did the work, it saved the result. Everything interesting was on your side of the screen.

That is changing, and not in the way most of the marketing says. The change is not that software got smarter at talking. It is that software can now be given tools — a file, an API, a phone line, another program — and be trusted with a small amount of judgment about which one to reach for.

What that is actually good for is an open question, and it is the question this place exists to work on. So we build the tools first. Every application on this page was built because something needed to be able to use it: a document that software can open and change, a workflow it can drive, a phone it can answer, a room it can work in (Fig. 2).

We are not a consultancy with a deck about this. We are a shop that keeps building the thing and watching what it does. Some of it works well enough to sell. Some of it is a machine in the corner making a noise we do not understand yet. Both are on this page, marked as what they are.

PeopleAgentsDocumentsread, write, drawWorkflowscall, branch, waitCanvasshare, edit, keepVoicecall, answer, routeThe officewhere the agents areA mailboxA phone lineSomebody's fileAn APIThe webOutside our machinesPeopleAgentsDocumentsread, write, drawWorkflowscall, branch, waitCanvasshare, edit, keepVoicecall, answer, routeThe officewhere the agents areA mailboxA phone lineSomebody's fileAn APIThe web
Fig. 2 — How work gets done here. People and agents are on the same rail because they open the same tools — the drawing would be a different and less honest claim if the agents hung off the side of it. The dashed lines are the boundary: below them the work has left our machines and is touching somebody’s real mailbox, file or phone line.

02

The Virtual Office

We gave the agents a building. It turns out that if software is going to work alongside people, it helps enormously for there to be a somewhere.

It is a 2D office you walk around. There are cubicles, two conference rooms, a kitchen, a terrace, a data centre and a rabbit. We work in it, and so do the AI staff: they research, write code, ship it, review each other’s work, and hand jobs back and forth with us.

They are not chatbots with job titles. Each one works with the same kit anybody here works with — documents and boards, the shared drive, meetings, a mailbox on this domain, the web, and for the ones whose job it is, the ability to write software and put it into production. When one of them works on a document it is the same document, in the same Canvas, under the same permissions.

The plan opposite is the real one (Fig. 3) — transcribed from the module the running service reads to build the room, not drawn for the website. Forty desks, because forty is how many places there are to sit.

It runs on our own machines and there is no public way in yet. What is on this page is what it is, not a demonstration of what it could be.

Enter the Virtual Office

The TerraceCar ParkRoadKitchenWCLookingGlassPisgahToxawayThe GymLobbyThe FloorBlue Ridge RoomDavidson RoomLoungeWaitingDataCentreThe Corner OfficeN010 tiles

Drag to pan

Fig. 3 — The Virtual Office, ground floor and grounds. Sixty tiles by thirty-four on a site of ninety-six by fifty-eight. Every rectangle here is transcribed from the office’s own floor plan module, which is the same data the running service ships to a browser to draw the room. Switch the live view on and the marks are the people and agents who are in the building, read from the office every few seconds; a mark slides between two readings so it can be followed, and both ends of that slide are real.

03

Things we build

Five of them, at three different stages. Nothing here is a plan — every one of these runs today. The stage says how far it is from a stranger.

Shipped
Go and use it now, at the address given.
In build
It works, and it is being changed underneath you.
Running here
We use it daily. There is no public way in.
  1. 01

    Docubend

    Shipped docubend.com

    Documents that software can operate. Convert, merge, split, sign, redact, read a scan back into text, draw a diagram and put it in the page. There is an API, so the thing doing all that does not have to be a person.

    docubend.com

  2. 02

    WebsNFlows

    In build websnflows.com

    Connect an API to an event to an action. It began as a way to drive Zoom from a script and is being rebuilt into the general case: a run that branches, waits, loops over a list, calls something else and does something with the answer.

    websnflows.com

  3. 03

    Canvas

    Shipped inside the workspace

    Documents, files, boards and diagrams that several people and several agents can be inside at once. Changes land live, every version is kept, and sharing is read-or-write per person. An account is needed to open it.

    sign in

  4. 04

    Voice & Communications

    Running here no public front door

    A phone system we built and run: a PBX, a numbering plan, an analogue gateway for the old handsets, a bridge out to the public network, and a voice agent on an extension that answers, transfers and takes messages. It is our own phone, not a product yet.

    how a call runs

  5. 05

    The Virtual Office

    Running here no public front door

    A building the team and its AI staff work in together. Everybody in it has a desk and the whole working kit — documents, boards, the shared drive, meetings, mail, the web, and for some of them the ability to ship code. It is the piece we are least sure about and most interested in.

    look inside

The voice work is the one that surprises people, so it gets a drawing (Fig. 4). We run the signalling — the part that introduces two browsers to each other — and then the audio goes straight between them and we are not in it.

Our serverBrowserBrowserSIGNALLING ONLYAUDIO AND VIDEOBrowserOur serverBrowserSIGNALLINGSIGNALLINGAUDIO AND VIDEO
Fig. 4 — Where a call actually goes. The dashed lines are signalling: we tell the two browsers how to find each other and then we are not in it. Where a network refuses a direct path, the call relays and we say so in the call.

04

Open questions

These are not features. They are the six things we have not worked out, which is a more useful list to publish and a much harder one to write.

  1. Agents

    How much can you hand to software before you stop being able to check it? We keep the working visible for exactly that reason, and we are still finding the line.

  2. Documents

    A document is still mostly a picture of text. What becomes possible when software can open one, understand its structure, change it, and hand it back without a person in the middle?

  3. Workflows

    Most automation breaks at the first thing it did not expect. What does a run look like if it is allowed to notice that, take another road, and say what it did?

  4. Communication

    A phone call is a very old interface and an extremely good one. What happens when the thing on the other end can look something up, do it, and tell you?

  5. Collaboration

    People and software in the same document at the same time. Presence, permission and history all have to mean the same thing for both, which is harder than it sounds and mostly unsolved.

  6. Signals

    Reading the outside world — prices, filings, news, a page that changed — well enough to be worth acting on. Our record here is honest: the operations work, the predictions did not.

05

Building it in public

A laboratory that nobody can see is just a room. The point of putting this on the open web is that the work becomes checkable — what got built, what shipped, what was tried and thrown away, and what the agents did while nobody was watching.

We are at the start of that. What is below is real: seven commits, with their real dates and the subject lines that were written at the time, from the repositories the work lives in. It is copied here by hand today. The next step is that it stops being copied.

Recent work, by date and repository
WebsNFlowsThe catalogue on the box is state, and the deploy was overwriting it
WebsNFlowsMerge and Loop: a run can come back together, and act once per item
DocubendMeasure the API, date the failures, and refuse to ship what nobody chose
DocubendRead a scan in the script it was printed in
The officeOperations: watch every AI employee working, and a switch that stops them
The officeTwo employees could not talk to each other, and nothing ever errored
The officeA silence became a plausible lie, which is worse

What this is not yet, and what it should be:

  • The log reads itself off the repositories instead of being copied here.
  • What the agents are working on, while they are working on it.
  • What we tried that did not work, kept next to what did.
  • A way for somebody outside to hand the office a job and watch it happen.

06

What interests us

Most software waits.

What interests us is the part that begins when it stops waiting — when a program can work out the next step, take it, and show you what it did.

Deciding whether that was any good is still a person’s job. We build for that, not around it.

A small team and a room full of software. If there is something you want built, or you just want to know what any of this can do, write to us.