Kybernetes

Not a chatbot. A shift that never ends.

Most assistants wait to be asked. Kybernetes runs whether I am there or not. Agents wake on a schedule, do their work, and either file it or escalate it. The interface is where I read what happened and decide what matters, not where I start everything myself.

It runs locally and is reachable only across my own network. Nothing about it is hosted, shared or exposed. That is a deliberate constraint rather than a limitation, and it is why the code stays private.

Crew

Every project, grouped by the playbook it runs on.

A board, a build cycle, a research loop, an inbox: four ways of running work, with the projects on each one under it. Each card says whether the work on it is idle or running right now, when it next fires, and how its recent runs went, as a strip of dots that fills in as they land.

Board

Columns, when columns are what the work needs.

One pipeline at a time, with each project in the stage it is actually in and carrying its open task count. A literal board, for when the honest answer is that I just need to see the stages.

Decisions

The only list I am required to read.

Each item says why it is being asked now, what is at stake, and what was already considered and rejected. One option carries the recommendation and the reasoning behind it. Everything an agent could settle alone has already been settled.

The morning brief

Arrives at 07:30 whether I asked or not.

What needs a decision, what ran overnight, a pulse line on the things I said mattered, the day’s calendar, and one flag. Written from what actually happened, not from what was planned.

Three kinds, and most of them stay quiet.

Talk to

Specialists I hand work to directly.

Each has a defined remit, so a question goes to the one built for it rather than to a generalist that flatters.

coordinator
engineer
researcher
reviewer
operator
business

Report in

Scheduled agents that produce something to read.

They arrive whether I asked or not, at a fixed time, which is the point.

daily brief 07:30
mail triage 07:45
scout sunday 09:00

Stay silent

Around eighteen agents watching health, cost, security and drift.

They are designed to produce nothing. When one speaks, something is wrong, and that is the whole signal.

health
cost
auditor
drift
ops

The three decisions I would defend.

It knows the difference between running and working.

An agent completing is not an agent doing something. Kybernetes tracks those as separate states, so a scheduled job that fired and produced nothing is visible as exactly that. Several agents carry explicit contracts against reporting work they did not do. That distinction came from being burned by the alternative.

It refuses to starve the thing I am looking at.

Background work is admitted against real memory available on the host, and a slot stays reserved for whatever is in front of me. An automation layer that makes the machine unusable has not automated anything.

It grades its own work with a different mind.

Output is reviewed by a model from a different family than the one that produced it. A model asked to check its own reasoning tends to agree with itself. Structuring the disagreement is the only version that has ever caught anything real.

The repository is private. You can ask.

Kybernetes runs against my own accounts, my own mail and my own network, so the source is not public and will not be. If you have a reason to read it, send one and I will consider it. Requests come to me directly and I grant access by hand.

Request access

I read every request myself. Expect a reply within a few days, or none if the reason does not land.

Every screen above is a capture of the running system against an isolated instance seeded with invented projects, an invented decision queue and an invented brief. None of it is my own state.