Mondrian

Giraffe's new interface is one large rectangle made of smaller rectangles, in theory infinitely nestable — the shape every complex system takes when it works.

Piet Mondrian spent a career reducing a canvas to rectangles — a few black lines, a few blocks of colour, every pane a different proportion, the whole thing held inside one outer frame. Our new interface is built the same way. One large rectangle, the screen, divided into smaller rectangles of different size and colour, and any rectangle can hold another composition inside it. It nests as far down as the work goes.

A screen of nested panes turns out to be the right picture of how modern work is actually assembled. An agent is a small computer program that does one thing. Give it a task and it runs; hand its output to the next; compose enough of them and you have a larger program — a workflow — that does real work. Each pane is one of those programs. The canvas is the super-program you build by arranging them.

An old idea, in many uniforms

This is one of the most durable ideas we have, and it keeps reappearing under different names. Alan Kay, who coined object-oriented programming, trained as a biologist and took his model straight from the cell. A body grows from a single cell into trillions of them, without a master controller and without falling apart, because each cell is a complete unit that guards its own state and speaks to its neighbours only through messages. Kay's regret was that people fixed on the objects; the big idea, he said, was the messaging — the stable interface between the parts.

Once you see the pattern you see it everywhere. Microservices: split a monolith into small services that talk over defined APIs. Subsidiarity: push every decision to the smallest unit that can hold it. Iran's mosaic defence: thirty self-contained provincial corps, each able to fight alone, a structure that shrugs off decapitation strikes because there is no single centre to remove. German Auftragstaktik — mission command — where the commander issues the intent and withholds the method, so each unit supplies its own and becomes a distinct element within the whole. Different centuries, different domains, one architecture.

Small self-contained units, autonomy inside, a stable interface between. That is the whole idea, and it is very old.

Decentralisation has a failure mode. Cells without a body are a tumour. What keeps composition coherent is the frame: the outer rectangle that bounds the inner ones, and the shared interface every part speaks. Mondrian never lets a block float free. Every rectangle registers against the same grid and is held by the same edge.

That is what the interface gives you. Each pane in Giraffe is a self-contained unit — a map, a model, an analysis, an agent doing a job — with a clean edge and a defined way of passing what it makes to the pane beside it. You build a workflow by arranging rectangles, and because any pane can hold another composition, the same move works at every scale, from a single measure to an entire masterplanning pipeline. The frame keeps it legible; the messaging keeps it composable.

Failure stays in its pane

A boundary contains more than layout. It contains failure. When a pane breaks, the break stops at its edge — the blast radius of a bad app is one rectangle, not the canvas. That single property changes what you are allowed to build. If a broken app could take down the workspace, every app has to be vetted like life-support. If it can only break itself, you can try things.

This is what makes experimentation cheap. You can vibe-code an app — describe what you want, let an agent build it, run it — and the worst case is a dead pane you close. Move the risk to where it is survivable and the pace changes: dozens of small, imperfect, purpose-built apps instead of a handful of heavy, over-engineered ones. The mosaic tolerates a missing tile. The monolith does not survive a missing function.

The economics turn on this. AI has made it cheap to build inside a pane — one job, a bounded context, code an agent can write and a person can still read. Aim the same tool at the whole system and the economics invert: the blast radius is unbounded, the failure modes can't be listed in advance, and a confident wrong answer with system-wide reach stops being a bug and becomes an existential risk. Contained, it is cheap; exposed to everything, it is catastrophic. Kept small, it pays — at the scale of a single pane you can afford to build bespoke, and to make it beautiful, because a mistake costs you one rectangle.

But containment on its own is sterile — apps that can't hurt each other can't help each other either. The other half is the stitching, and that is the foundation Giraffe provides: the stable interface the panes pass data across, and the orchestrator that arranges them into a workflow none of them performs alone. Isolation on one side, composition on the other. Let the parts fail small so the whole can grow large.

This makes Giraffe two businesses at once. One is a solutions business — each solution an app, or a series of apps, built for a specific problem a developer or a planning team actually has. The other is the foundation beneath them: the platform that makes those apps safe to run, embeds them where the work already happens, gives them the geometry and data to act on, and orchestrates them into something none of them does alone. The solutions are where the value is visible. The foundation is what lets them be built fast, fail small, and still add up.

The future of this work is fewer people each conducting more programs — arranging agents the way Mondrian arranged colour, until the panes together do the work. The canvas is the point. Mondrian is where you paint it.