Our own agency
We build together. Golden Ratio is a small company of agents and humans, each making the other better at what it does. Work arrives as events and every piece of it carries an intent; agents know what to do, where to put the work, and — the hardest part — when to ask; people decide what matters and stay answerable. We run that arrangement on ourselves first, in public, and it is the same one we install in your work rather than describe in a deck.
The problem
Every company now says the machines do the work. In most of them the machine is a chat window: somebody pastes the context in, copies the answer out, and holds the whole thing together with memory and goodwill. The work has no shape anyone else can see, and nothing moves unless a person pushes it. We wanted the opposite — a company where every piece of work carries its own declared intent, where the work moves on its own, and where no agent may act outside what it was given.
The wrong turns
We began by building all the orchestration we could imagine, which is to say too much of it. The early versions re-triggered themselves: one piece of work could generate hundreds of events in an hour, and the company spent a morning talking to itself while nothing was delivered. We stopped the machinery on purpose and left it stopped until we understood it. Two smaller versions of the same mistake are worth naming, because we still make them: two writers updating one record, which is how a company ends up with two truths, and a decision changed in one place while three other places still enforced the old one.
What worked
Work arrives as an event, and every piece of work declares what it is for before anything runs. It is claimed only when what it depends on is finished; one writer owns each record, so two parts of the company never disagree; and the decisions a person has to make are explicit points where the work stops and waits rather than drifts. Nothing is called finished because somebody said so: verified means read back from the record, and the evidence is the change itself rather than a summary of it. The machines do the long part. The humans decide what leaves the building.
What it costs
Autonomy is not free, and the price is paid in the currency the brochures never mention: time, and attention at the gates. Every piece of work passes several approval points, and each one needs a person to read something properly. A decision that changes has to be carried into every place that enforced it, and when it is not, the old rule wins quietly. A company like this is slower than one with no rules, and much faster than one whose rules live in people's heads. That is the trade, and it has to be made deliberately rather than discovered.
What it would look like for a client
The same shape travels. A piece of work with a declared intent, one writer at a time, approval points where a person decides, everything written down, and the finished change as the receipt. We are not selling the machinery; we are offering to run your process the way we run ours — the machines doing the long part, the humans keeping them honest, and every promise traceable to something you can read. If your work is a queue of small, well-scoped pieces, this is where it pays first.
The short version of the design: nothing here is a folder of instructions. Every piece of work is an event with an owner, a scope and a stated purpose, and the company moves because events move it. An agent that has been given one piece of work cannot decide to do a neighbouring one, and cannot quietly widen its own remit — which sounds restrictive until you have watched what an unbounded assistant does to a codebase.
The rules that matter most are the boring ones. One writer owns each record. A piece of work is claimed only when everything it depends on is finished. Every decision a person has to make is a visible point where the work waits, so nothing stalls invisibly and nothing proceeds on a guess. And "done" is never a claim: it is read back from the record, and the change itself is the evidence.
What that buys is not speed in the ordinary sense. It is that the company can be left alone — for an afternoon, or overnight — and be in a state somebody can explain in the morning. That is the part clients ask about first, because it is the part their own AI projects are missing.
We are telling this one with the wrong turns left in, and it is unfinished on purpose. The next entry in this section is what the same routine looks like when it runs inside somebody else's repository.