TetrAgora users pick a site and input needs into a spatial planning algorithm that configures modular parts. Quadruped robots build it, learning reversible assembly. As needs change, only affected parts are reconfigured, so the architecture adapts.

Site packing grows rooms across a plot, factoring in setbacks and envelope bounds. Adjacencies govern touching walls as rooms fill the site. Testing multiple random seeds, the packer scores each run for brief match, compactness, and left over space.

Each part's form is derived, not drawn: from the system grid, it passes through aggregation, merging, surface negotiation and refinement. Every curve traces to a rule; the bone-like character is the residue of maximising contact between neighbours.

The latest prototype. Four degrees of freedom per leg: the hip yaws, the knee, ankle and toe pitch. A release-and-attach mechanism at the foot means the same limb that carries the robot also carries the piece.

At model scale, magnets stand in for the connection. Polarity does the work: aligned, the foot holds; opposed, it lets go. A placeholder joint, built to test one behaviour, attaching and releasing with no latch.

Windows aren't cut from finished walls; they're placed first. Each blue plane is a protected zone: agents build up to it, never through it. Fenestration constrains growth instead of subtracting mass. Panels log fill, 800 builders, wandering at zero.