ARCHIMUS

What happens when an AI agent is given the tools to create, the knowledge to critique, and the freedom to improve itself? We explore this question through Archi, an experimental agent that learns about architecture through iterative creation and self-critique.

The Experiment

Archi begins with:

Access to 3D modelling tools
A corpus of architectural knowledge
The ability to update its own skills based on what it learns

The agent is then set loose to explore the domain of architecture, creating designs, evaluating them against learned principles, and refining its approach.

Self-Improvement Through Skills

A key feature of Archi is its ability to create and modify its own skills. As it learns what makes good architecture, it encodes this knowledge into reusable capabilities:

proportion-check: Evaluate whether proportions follow classical principles
light-analysis: Assess how natural light interacts with a design
flow-evaluation: Analyze movement patterns through spaces

Emergent Understanding

Over time, Archi develops increasingly sophisticated understanding:

Early Iterations

Initial designs focus on basic structural validity—walls connect, roofs cover spaces, doors lead somewhere.

Middle Period

The agent begins to consider human factors—sightlines, circulation, the relationship between public and private spaces.

Later Work

Designs show consideration of context, materials, and the emotional qualities of space.