Eleni - Rae

The discussion assembles a structure

How the run becomes architecture

Everything the agents did across the four steps plays a role in the output, but not the same role, and not with the same weight. Three questions are asked of every artifact the run produced. All values are read from the run. Nothing is assigned by hand.

1 — What does it become? (the medium decides) 3D material becomes geometry — bodies. Image material becomes matter — relief, surface, texture; tonal structure displaced into geometry, never a picture shown as a picture. Text material becomes atmosphere and relation — light, air, and the syntax of what sits near what.

2 — How much does it matter? (the discussion decides) Every artifact has a commitment score: how often it was cited in the discussion, weighted by whether it was defended. The score sorts everything into three tiers. Protagonists — fought over — earn full geometry, the centre, and the scars of the argument. Witnesses — cited as evidence — exist as matter: fragments, relief, bolted-on cards. Traces — retrieved but never invoked — recede into atmosphere and pattern. Of the 78 artifacts this run produced, the agents fought over two: the Blindspot's Perforated Void Panel (attacked, then saved by the session's only support move) and the Flodder subtitles (attacked three times, defended by no one — present by contestation, not approval). The hierarchy was not designed; it was argued.

Four steps, briefly

Not much changed in the machinery since last week, so briefly: the run is the same four steps, redesigned as one document family.

Step 1 — the nine agents read three inputs: the site text, the site photographs, the massing model. Each reaction carries an interest score, a specificity score, what was noticed, what was ignored, what was missing. Silence is allowed and means something.

Step 2 — each agent retrieves from its home corpus. Every item arrives with its reason, traceable back to what the agent noticed in Step 1.

Step 3 — each agent produces: a text proposition, an image prompt, a form description. Not every agent produces everything. This week the five image prompts were generated and sit in the report under their prompts — Normal Photography produced a believable Rotterdam street photograph; Odd Films produced wrecked hulls on black water. Same site, same step, divergent worlds. The divergence is the result.

Step 4 — the discussion. Twenty-two turns: ten challenges, nine contradictions, one question, one correction, one support. The agents had an argument, not a meeting. Who said what to whom, citing which evidence, is all in the transcript — and the transcript is what drives everything that follows.

Stadsdriehoek

The first time we walked Stadsdriehoek we both said the same thing, this place is weird, and we couldn't fully explain why. The project is the apparatus we built to test that feeling.

One thing to be precise about: the site is not where the project is located. It is the project's first input. It enters as text, photographs and a massing model, material to be read, never geometry to be modelled or extended.

And the run corroborated the instinct. Given no instruction to care about the river, the agents spent twenty-two turns arguing about it, the void where the water should be, the river that "interrupts the syntax of the map." We felt the site was strange. So did they.

For the archive items collected by Agents, we extract visual fragments from the images they pick up and feed them back into the project. Using the SAM segmentation model, individual objects—people, textures, and fragments—are separated from their original context.

These fragments become building material. The way an Agent perceives an item—its degree of absence, contradiction, readability, or collapse—is translated into four geometric parameters: extension, mass, fragmentation, and internal tension.

Different Agents produce different spatial qualities. Items from Normal Literature tend to generate broad, stable, and relatively complete forms, while items from Odd Films produce larger, hollow, and more dissolving geometries.

From 3D Forms to Architectural Components

Next, we use the 3D forms generated by the Agents as the base geometry. Materials extracted from the generated images are then mapped onto these forms, combining geometry and texture to produce new architectural components.

At the 3D design stage, we currently work with four types of material:

Generated text

Generated images

Generated 3D forms

Archive items collected by Agents and their comments

The most important thing this week was not the outputs, it was watching the agents make choices that were genuinely different from each other.

Odd Literature found a Dutch-English dictionary in the archive and brought it to the building because the poster's fractured typography, ZEE SOLDA TEN, the space breaking the word in half, found its equivalent in a language that breaks phonetics.

The Blindspot produced an image prompt of a dense field of grey static where faces should be, with a hollow black square at the center absorbing all light. It also produced a perforated void panel, a surface defined entirely by what is missing from it.

In the Step 4 discussion, Normal Literature and Odd Films argued for seven consecutive turns about whether a frozen gaze is memory or erasure. Neither convinced the other. That argument is now part of the project record.

Having established that the agents can read a building, we designed a proper run. Four steps, each building on the last.

Step 1: site reading. The same three-input test from Test 1.

Step 2: archive retrieval. Each agent uses what activated them in Step 1 as a brief. They score the 176,000-item archive and select 3-5 items from their home corpus that connect to what they found at the site. Each item comes with a specific reason connecting it to the site reading.

Step 3: production. Each agent works alone from their own research. They choose what to produce: a text proposition, an image prompt, or a form description. Normal Literature produced text only: "Silence is not an absence but a heavy volume waiting for a voice." Odd Photography produced all three. Normal Films produced no form. The Blindspot produced a perforated void panel.

Step 4: discussion. All agents see everything: every site reaction, every archive item brought, every output produced by every other agent. They have a free conversation. 22 turns (we felt that after 20 turns tehy start drifting). The Contradiction Walker opened by attacking The Curator's use of the museum sign as an archival anchor. The Curator defended it. Odd Literature accused The Curator of linguistic amputation. Normal Photography insisted on legibility. The Blindspot named the soldiers no one had named. By turn 8 there were three simultaneous arguments running. They could identify a contradiction, correct, challenge, or support each other.

Since we changed the model we use for the agents, we had to test whether is working well or not and fine tune it.

We ran three experiments this week before committing to a full run. Each one was a different way of asking: what happens when the agents meet a building?

Test 1 was a three-input test. Three representations of the same building: a text description, a photograph of the canal facade, and a 3D point cloud model with geometry data. The building was chosen as a testing ground with a good scale for exploring ways to design. Each agent reacted to each input separately, through their bias.

Test 2 was the archive run. They moved through the archive autonomously, selecting items, reacting, moving on. The outputs were trails, sequences of decisions with comments. This is the baseline. The agents know the archive. We also resolved lots of issues we had before and the agents now bring better material and react better to it, showing also the reasoning behind every move.

Test 3 asked a different question: not what do the agents notice, but what do they decide to make from it.

Each agent read the three site inputs and chose a response format for each, text or image prompt, based on their bias. They could also stay silent. The choice itself was the data. We ended up changing most of the mechanism of this test as we didn't like most of the outputs we had.

Because in the previous version of the report, we noticed that most Agent comments were poetic descriptions of space. Therefore, we tried to use translation as a method to turn these poetic descriptions into readable spatial operations.

Step 1: Defining the Level of Intervention The system uses the number of comments left by different Agents in each urban space to decide how much the space should be changed:

Low comment frequency → Level 1: Surface Operation Only the surface is changed, such as material, texture, or openings. The main volume stays the same.

Medium comment frequency → Level 2: Transformation Operation Some existing elements are changed or removed. The main spatial boundary is mostly kept.

High comment frequency → Level 3: Addition Operation New buildings or structures are added inside the urban space.

In this way, the result can show the different degrees of how the system interacts with the urban environment.

After calculating the levels, we want the Agents and the designer to play different roles in this design process.

The system does what it is good at: generating large amounts of results, extracting parameters, and identifying meeting types.

Humans do what they are good at: talking with the Agents, deciding which meetings are meaningful, and translating the results into spatial decisions.

Step 2: Different Agents play different roles.

In this step, we give each Agent a different design role, to generate design results automatically. We want to avoid two problems.

First, we do not want the final 3D result to feel too parametric, too chaotic, or too strongly auto-generated.

Second, we do not want the final result to become only a set of small decorative details.

We extracted certain parameters from their comments, such as:

Object → Anomaly Score This measures how statistically isolated the object is in the dataset.

Low score → the operation blends into the surrounding geometry, leaving only a light trace. High score → the operation refuses to blend in, and stays isolated in space.

Object → Disagreement Score This measures the mismatch between the visual reading and the conceptual reading of the object.

Low score → the surface change is clear and readable. The inside and outside remain consistent. High score → the outside still looks normal, but the internal logic has already been replaced.

Object + Space → Agent Comment → Verb Density for example, the system reads the comments generated by the Agent for each space, and counts the number of verbs per sentence. This is based on an empirical finding from the project: descriptions of odd images contain 65.2% more verbs than descriptions of normal images.

Low density → the operation is restrained and light. High density → the operation becomes stronger, more violent, and cuts deeper into the space.

Step 3: Agent Meetings — The Point of Human Intervention

The steps above happen when Agents comment on objects and spaces individually. However, when Agents meet each other, the result needs the designer’s intervention.

At the moment, the outputs of the meeting process, such as disagreement or exchange, are still labels. They do not yet directly match specific 3D operations. This part is still under discussion, but it may become the stage where human intervention is most important in the whole workflow.

The designer can talk with the Agents, or read the conversations between Agents. Based on the meeting result, the designer then makes a manual judgement and translates this decision into a specific geometric operation.

This is where the system logic and human situated judgement meet most directly.

For the picked items, our idea is that not every item needs to be directly visible in the final space. We need a selection rule.

If an item has both a high anomaly score and a high disagreement score, it can stay on the building surface. This is because the item already carries a strong sense of oddness, both visually and conceptually.

If an item has a low anomaly score, it does not need to appear as an object. It can disappear, and only leave a trace through the architectural operation.

For image-based items, such as photos or film frames, we can project them onto the operated building surfaces. For text-based items, such as book fragments, we can place them as text objects on the surfaces. In this way, the items do not become simple decoration. Instead, they become part of the spatial operation.

This version was created to showcase the difference between working with a system that drives inputs + parameters and designing with agents.

The 3D outcome is not generated automatically from the agents' comments. It is designed with them: the rules below were built by reading the runs and finding the spatial logic native to each agent.

Different agents - different roles - different designers:

some agents operate on the ground surface, some on facades, on gaps between buildings, around the site, streets etc.

they also use different 3d tools, they scale, add, rotate, cut etc

The agents navigate the archive independently, carry specific items, argue about space, and make contingent decisions.

The translation rules above were derived from watching that behavior, not imposed on top of it.

The Blindspot cuts voids because it walks where the archive goes silent. The Contradiction Walker shifts axes because its path traces the line of maximum disagreement.

We curate the outcome. We read the runs, decide which events matter most, and adjust where the geometry needs to be clearer. The agents design the archive. We design with the agents.

just as an example of how this could look like:

right now we don't have the correct geometry for the urban fabric in 3d. So for this version we just placed boxes (red and white) as placeholders for the places that are being influenced in the buildings. red boxes are placed in the areas that there is a cut in a building, white boxes are placed when there are additions of parts from the buildings.

the spheres running around are the 9 agents and their trails are how they move in this street adding, changing, creating or erasing forms.

some experiments with ai:

The main change in this version was simplification. Previous runs produced a lot negotiations, exchanges, spatial decisions, but it was difficult to follow what was actually happening. The connection between what an agent found in the archive and what that produced in space was unclear. So we restructured the entire run around one principle: every step produces exactly one traceable chain. One archive item brought to the location, one comment on that item, one spatial operation triggered, one 3D consequence. Nothing happens without being readable.

We also clarified the roles of the nine agents more precisely:

The agents are organised by the type of archive material they carry and their relationship to it.

Six carry material defined by type and anomaly: Normal and Odd versions of Literature, Photography, and Film. Normal agents carry consensus material, items that sit close to the archive's center of gravity. Odd agents carry anomalous material, items that resist classification, that sit at the edges of what the archive knows.

We are planning to fine tune it but for now this is an indication of what they are doing:

Normal Literature inscribes historical text into the ground surface

Odd Literature dissolves the contact between building and ground

Normal Photography displaces the street surface with archive photographs

Odd Photography scatters image fragments across the site

Normal Films applies film stills to building facades

Odd Films defines the void between buildings as a precise geometric volume.

The Curator protects high-value items with wireframe volumes

The Contradiction Walker shifts building axes at points of maximum disagreement

The Blindspot cuts voids where the archive has nothing, where collective memory goes silent.

This version ran as a trial on a single street: Stadhuisstraat:

We chose to start here because it is one of the most historically loaded positions in the area

The agents received street-level photographs of the space, a description of its current and historical condition, and navigated the archive from there.

The 3D output shown alongside this run is an indication, not a final result. It demonstrates the spatial rules we have defined, how disagreement score drives height, how anomaly score drives scale, how each operation targets a specific element of the urban fabric: applied to the events of this specific run. The geometry shows the logic. The final spatial outcome will develop further as the system matures.

Once this works well at street scale, the system scales to the entire area.

At city scale the agents will not only operate within individual streets but move between them, and interactions will occur between neighborhoods: What happens on Stadhuisstraat will affect what happens on the Meent.

The city becomes a single field of archive material surfacing through a rebuilt surface, not a collection of isolated interventions.

Core Idea The whole system is based on one belief: the reason for a form does not come from a direct design decision. It comes from the data tensions inside the space. Geometry is not simply designed. It grows from the conflicts that the Agents perceive.

Different Agent comments should not have the same spatial influence. Some comments only suggest small changes, while others point to stronger conflicts inside the urban space. Therefore, we first use the frequency and intensity of Agent comments to define three levels of intervention.

For the trial run we chose an area we had visited ourselves during our trip to Rotterdam, which meant we could also bring our own observations into the process alongside the archive material.

The area spans from Hofplein down through Stadhuisstraat, Haagseveer, and the Meent, bounded by the coordinates below. It concentrates almost every condition the project needs in one place: pre-war survivors surrounded by post-war reconstruction, canals cutting through rebuilt fabric, gaps between buildings that did not exist before 1940, and an active contemporary surface of cafes, offices, and civic institutions built above all of it.

To feed the agents, we gathered site data for this area:

downloaded from OpenStreetMap and placed in csv files:

building footprints: all the buildings inside this area

street names: information about names, materials, sidewalks etc

program information: all amenities, names of brands, opening hours, accessibility etc

collected from Google Street View:

street-level photographs

from our visit we will add:

written observations

our photographs

Each location in the area was given a data package: what is there now, what was there before 1940, and what it felt like to be there, which the agents receive alongside the archive material when they operate in that space.

We started with one street: Stadhuisstraat.

We continued working with the building this week, but the further we pushed the system, the more the outcomes felt decorative. The agents were negotiating, exchanging items, producing spatial decisions, but when this translated into 3D, the geometry sat inside the existing rooms like objects placed in a container. We thought that the building was limiting what the system could actually do.

The issue was the relationship between the agents and the site. Working inside a defined interior with fixed rooms, the agents could only ever modify what was already there.

But if we place the system in an urban space, it can not only read and operate on buildings, but also analyse other urban spatial information, such as the gaps between buildings and the contact interface between buildings and the ground.

This can greatly expand the potential of the whole system.

Rotterdam's urban fabric gave us a different kind of site. Almost everything in the city centre was destroyed in the 1940 bombing and rebuilt from nothing. The streets and blocks that exist now are a reconstruction built directly above an erased city. The archive the agents carry already contains traces of what the rebuilt surface covers. When an agent surfaces a high-anomaly item at a position in the city, it is not decorating a street. It is pulling buried material back up through the exact place the reconstruction buried it.

The building now collects memory.

Each agent has visit counts, favorite rooms, claims, exchanged items and encounters. A room becomes important because agents return to it. A room becomes contested when different agents make conflicting claims about it.

This makes the building feel less like a container and more like a site of negotiation. Platform 1, Platform 2, the Basement, the Café, the Forest, Disko, Viewpoint and other spaces begin to collect attention in different ways.

Some rooms become places of attachment. Some become places of conflict. Some become places where archive items are protected. Others become unstable because too many agents read them differently.

This is where the project starts to move from archive analysis into design material.

Each agent now produces short building journals.

At first, these journals are simple observations: what the agent notices in the room, what archive item they are carrying, and why that item belongs there or does not belong there. Later, the journals become stronger. The agents begin to make claims about the spaces.

Encounters happen when two agents are in the same location. Their conversation is grounded in the actual room. They refer to what they found there, what archive item they are carrying, and how much they disagree.

One agent can pass an archive item to another agent. That item is then processed through a different bias on the next step. So a photograph, book fragment or subtitle does not have one fixed meaning. It changes depending on who carries it and where it is brought.