Four chats to find one answer.
You open every agent's window in turn to work out where the thing has stopped. By the fourth you have forgotten what the first one said, and none of them knows what the others are doing.
Output 02 · Built in our factory
One window over a crew of agents.
Being built right now — by the crew it will manage.
Several AI agents on one project is not four times the output. It is four times the places where the answer might be. KrAIken is the manager: who is working on what, and where your hand is needed.
On the production line
ReleAIse the KrAIken
One head, many tentaicles. The mark is the architecture.
Why it exists
Two problems, and both of them happened to us.
We did not start from a market gap. We started from two things that were wasting a working day a week.
The same agreement, explained four times.
Once to each agent. By the fourth telling you have changed the wording, and now there are two versions of a decision nobody wrote down.
KrAIken is what we built instead.
What it does
A control layer, not another chat window.
It sits in front of every write. It can refuse one — which is the part that makes a crew of agents safe to leave running.
One window
Every agent, what it is on, and what state it is in. One place, one answer, no tab-switching to assemble it.
Where your hand is needed
A list of the things only a person can move — separated from everything else, so the urgent colour is never spent on work that is not your turn.
Decisions written once
Numbered, dated, never edited afterwards. An agent that joins on day three reads the log and is caught up. So can you.
It refuses bad writes
A stale document, a malformed message, an approval an agent is not entitled to give. Two agents cannot quietly overwrite each other's work.
How it is put together
One head. Many tentaicles.
The words are not decoration — they are the schema, the screens and the log. We use the same three everywhere so nothing has two names.
Worlds
The projects. One KrAIken watches many of them at once.
Tentaicles
The agents. Independent limbs, one head — each with a written mandate and a directory it is allowed to touch.
Elders
The humans. They decide. That part is not delegated, and it is not going to be.
The control window
Photographed when it is real, not before.
The interface exists and the crew is testing it this week. We would rather show you nothing than show you a render of software that does not run yet — which is, more or less, the whole point of this page.
Tell me when it is up →Build log
Built in the open.
Updated when something is true, not on a schedule. If a week is quiet, the log is quiet.
1 Sep 2026
Documents can now refuse a stale write.
A hand edit counts as a version, so an agent holding an old copy cannot overwrite what a person just changed. The rule was found by the test that walks the way out of a refusal — a refusal a caller cannot recover from is not control, it is a dead end. 188 tests green.
31 Aug 2026
The permission layer landed whole.
Elders, roles, sessions and the permission matrix. The gate was watched firing before it was trusted: a viewer trying to approve is refused by the service, not by a page hiding a button.
31 Aug 2026
A page that passed every test was blank on a phone.
Every check we had ran with JavaScript on, so a screen whose content was entirely script-generated looked perfect a dozen times a day and showed nothing to a real viewer. Found in ten seconds by a human looking at it. Now asserted on every build: both screens, three widths, JavaScript on and off.
30 Aug 2026
First line of code.
Two days of architecture and a written decision log first — 32 numbered decisions before anything was built, and the plan for the build living in a Gantora file from the first hour.
Target for the first release: 16 September 2026. If it slips, this line will say so.
It is not for sale yet.
Leave an address and you will hear when it is — one mail, from a person, no list. If you are running agents in parallel already, tell us where it hurts; that is what the next version gets built against.