Autumn Show 2026
unit-code
Nine Acts of Ostranenie takes the post-war reconstruction city as its ground, operating where the uncanny is a structural condition built into the erasure and replacement of the urban fabric. Through a vast archive of books, photographs, and film, nine AI agents navigate this territory to surface divergent readings of the city, intentionally conflicting with both each other and the machine's statistical logic. This disagreement serves as the generative ground for design, shifting the focus toward the friction between interpretations instead of a singular brief. At each site, a proposal unfolds across three registers, text, image, and built form, acting as a scaffold that carries these internal tensions forward rather than erasing them. By translating these complex echoes into the landscape, the project moves beyond prescriptive planning. The city's strangeness is neither hidden nor ignored; instead, it is given physical form through an architecture capable of holding, navigating, and negotiating the realities of a contradictory urban environment.
The familiar has become unreliable. Ostranenie interrupts that habit. Every act of seeing is partial. Oddness depends on expectation, context, interpretation and prior knowledge. The project asks a LLM to work in the opposite direction.
Rotterdam became a ground for exploring oddness because the city can be encountered in many versions. Post-war reconstruction produced replacement and discontinuity, while literature, photography and film add another layer of instability.
Literature holds Rotterdam as written, described and imagined. Photography follows ordinary moments not arranged for the camera. Film records the city through movement, duration and action, connecting what it looks like with what happens within it.
The unified map is a data-driven archive of all scraped items. A search retrieves works across all three datasets by anomaly score, while cross-dataset links reveal patterns and contrasts. The most unusual items surface first.
The agents share a language model, but they are not given the same position within the project. Difference is therefore not produced by asking one model to imitate nine personalities. It is built into what each agent can access and what it notices.
The site is a river-edge fragment of Rotterdam, shaped by reconstruction, traffic, memory and everyday use. Each agent receives the same three forms of evidence: a written reading of the site, site photographs and a three-dimensional model.
"Site inputs A+B+C: image · location · viewpoint · conditions.
Reactions: nine agents read the same site differently.
Retrievals: reactions open archive; fragments return.
Productions: fragments become text · image · three-dimensional propositions."
"4. Discussion
agents meet through their productions
they compare · defend · challenge · delete · alter · preserve"
At every location, participating agents decide which form—text, image or three-dimensional space—best expresses what they want to say. Each proposal is designed only by the agents who chose to participate in that particular encounter.
All reactions, retrievals, productions and discussions are mapped across fourteen site locations using the coordinates of their original inputs. Circle size indicates agent participation; the largest marks the location shared by all nine agents.
Through the mcp connection, agents link to Blender. They read the scene and observe others. They create, move, scale, deform or remove forms, apply materials, and control cameras and lighting; each action becomes visible and available to the rest.
The agents were asked what activities they wanted to host, how many spaces each required and what dimensions they should have. Their responses were translated into fourteen initial volumes, then configured through shared topics and disagreement.
Through the mcp, each agent created a camera inside its spaces, producing a live feedback loop between seeing and making. The agents did not design the space as an external object; they shaped it from the position of its future occupant.
The agents began constructing and positioning components according to the needs of each activity. Every screen, divider, frame, grille and opening was introduced to shape how the space is occupied, perceived and navigated.
The agents leave their own spaces and enter those of others. Each carries one of the forms produced during the earlier run and offers it as a gift, introducing a fragment of its own position into another agent’s spatial argument.
Each agent selects materials according to the metadata, descriptions and material associations contained within the archive it was trained on. These choices translate accumulated knowledge into colour, texture, transparency, weight and atmosphere.
The same process was applied across the remaining locations. At each, participating agents reacted, retrieved archive material and selected the output type that best expressed their position. The site map holds these situated interventions together.