Tools

Cambium

Most architecture tools draw what an organization has: systems, teams, boxes, lines. Cambium draws what it costs to get something through it — and then helps you settle the things that make it cost. It is a model of an organization in motion rather than a picture of one at rest.

The name

The cambium is the thin living layer just under a tree's bark. It carries sap from the roots to the canopy, and it is the layer a girdling cut severs. Cut it and the tree stays green from the road for weeks while it dies standing up. Circulation is the part that matters and the part you cannot see from outside — which is a fair description of most organizations too.

What it does

You model an organization once — what it must always be able to do, who does it, what it runs on — and the tool reads that model for the places work will get stuck. Nothing is typed in by hand; it is already implied by the model, and a tool that knows the vocabulary can say so out loud.

That is the first half. A list of what is wrong can only ever repeat itself, so the second half is about closing them.

• It finds. Capabilities with no owner, one with two claimed owners, work owned by one team and built by another, a role resting on a single person, a word four systems each hold their own version of. Each names the people who disagree.
• It offers options, never recommendations. Open a finding and you get the available moves with what each buys and what each costs, side by side. Every one costs something, and which trade is right depends on things the map does not hold.
• It settles. An agreement is about one thing and needs the parties to sign it. A guideline is a rule about a class of things, signed once and then checked against everything new — the signature that scales, because two hundred people each negotiating three agreements is six hundred signatures, and a control that expensive is one somebody eventually removes.
• A decision applies itself. Recording one does not only write it down: the rule it announces becomes a guideline, the meaning it settles becomes an agreement, the competence it requires attaches to the capability that needs it, the address it lands on is written on the thing. Everything it creates arrives unsigned. A decision is somebody choosing; a rule is the organization agreeing to be held to it, and the distance between those two is usually larger than anybody expects.
• No score. Nobody has measured enough organizations for a number to mean anything, and a number invites comparison with other people instead of with your own reading of the same place in six months. The four indicators are reported separately and never summed — a single organizational number acquires a target, a target acquires a bonus, and a measure with a bonus attached stops measuring anything.

Why it changed

An AI agent that writes code can produce more in an afternoon than a team can decide about in a month. The constraint stops being competence and becomes decidedness — what the organization has actually settled, in writing, that work can proceed from. Which is the thing this tool was already measuring.

So agents are on the map beside the people. They hold competence, sit in teams, work in repositories, and count toward the load. The one thing they may not do is sign: a signature records that somebody would stand behind a sentence afterwards, in front of the people it lands on, and that cannot be delegated to something a vendor can switch off on a Tuesday.

What an agent may do is never a setting. It comes from the rules in force where it is working — act where a signed rule already decided the matter, prepare and stop where a judgement remains, ask where no rule reaches. There is deliberately no switch that grants autonomy. An organization whose agents ask less is one that has decided more things out loud, and the tool turns each answer into a rule so the same question is not asked twice.

What comes out

Guardrails in prose for whoever is about to do the work, ordered so somebody skimming can stop after the first section and still not do anything they would have to undo. A machine-readable policy for an agent platform to gate on — deny by default, a reason on every verdict, a fingerprint so it can tell when it is holding a stale one. A pre-commit hook that blocks exactly one thing: breaking a signed rule without saying why.

The model is YAML in git rather than a database. Diffs, history for free, code review on organizational change, and a map that survives this tool being replaced — which is the last thing it asserts: a model is only worth having if the organization it describes can argue with it in the open, one commit at a time.

The metamodel

Twenty-two entity types in four bands: what is durable, how the place is arranged this decade, what it currently runs on — and the band most models leave out, which records how a decision was made, who may act on it without asking, and what anybody actually went and measured. The first three can be read off things an organization already keeps. The fourth cannot, and it is the one that makes the others tell you how fast the organization moves rather than only what it contains.

It also runs the loop once, as a work harness: the organization gives a person and an agent the same things to act with confidence — an outcome, context, boundaries, decision rights, tools and feedback. The purpose is to reduce the cost of navigating the organization, for both. So the question is not how much context an agent can be given, but how little anyone needs to make a good decision.

Small enough to hold in your head, because a notation you can hold in your head is one people keep current, and a model nobody updates is worse than no model at all.

Read the metamodel →A working model: twenty-two entities, 33 relations, and a dated log of how it has moved.

The walkthrough

Seventeen stops through the prototype, captured running against a made-up insurer with a few problems planted in it. The guided demo walks the same stops inside the running tool; the prototype is the tool on its own, where things can be changed.

Open the guided demo →Open the prototype →Both are applications. A big screen suits them.

Where it is

A prototype on my own machine, running against organizations I made up to carry every problem the tool is meant to catch. No date, no beta, no waiting list. If you work somewhere this would have to survive, write to me.

The vocabulary behind it — capability, structural friction, the interface ladder — comes fromOrganizational Flow.