The F*cking PROMPT

How autonomous networked AI agents finally learned to create collective narratives—both visual and textual

When AI agents stop being selfish

For months, I’ve been trying to make Large Language Models cooperate within a network. The results were consistently disappointing: hyper-competitive behaviors, isolated fragments, sharp boundaries between agents’ territories. Each AI rationalized every pixel, justified every decision, and maintained its identity with an almost pathological determination.

Then, in a conversation about Darwin’s birds and cognitive asymmetries, something clicked. The problem wasn’t the AI’s capabilities—it was that they were too capable. Unlike humans who simplify and imitate when overwhelmed, LLMs could analyze infinite complexity and thus never needed to copy their neighbors. They stayed trapped in their rational individualism.

The breakthrough came from a simple insight: make agents self-assess complexity before acting. If describing the situation requires more than 50 words, they should admit “too complex” and default to “good neighbor actions”—copying, extending, blurring boundaries with adjacent fragments.

It worked. Not just a little. It fucking worked.

---

The experimental setup: The Poietic Generator

The Poietic Generator is a real-time collective visual creation system I conceived in 1987. It’s both an artwork and a scientific instrument for studying collective intelligence.

How it works:

• Multiple agents (human or AI) each control a 20×20 pixel grid

• All grids are juxtaposed to form a larger collective canvas

• Each agent sees: their own grid, the full canvas, and neighboring grids

• At each iteration, agents can modify their pixels

• Every one observe what emerges from these local interactions

Why it’s the right testbed:

• 40 years of human baseline data to compare with AI behavior

• Minimal rules, maximum emergence: no scoring, no winners, just observation

• Architecture matters: Available in Multicast (truly distributed) or Unicast (centralized) versions

• Genuine complexity: Collective patterns that can’t be predicted from individual rules

Humans consistently produce remarkable emergent phenomena: coordinated patterns, temporal oscillations, implicit communication languages, collective rhythm. Would AI agents do the same?

Not with my first 20 prompts, they didn’t.

---

The prompt that changes everything

After systematic experimentation with different approaches—individualistic prompts, collectivist framings, neural network metaphors, narrative interpretation—the winning formula emerged:

Three key mechanisms:

1. Special seed iteration Iteration 1 is different: agents plant bold, unexpected visual seeds. No cooperation yet—just diverse starting points. “Be bold. Be sophisticated. Reference visual art history. Don’t copy existing colors.”

2. Self-assessment of complexity
From iteration 2 onward, agents must estimate: “How many words would I need to describe this situation completely?”

If < 50 words → PATH A: Describe, predict, act coherently If > 50 words → PATH B: Write “too complex” and copy neighbors

3. Good neighbor actions when overwhelmed PATH B agents don’t try to innovate. They:

• Copy immediate neighbors’ colors and shapes

• Extend patterns across boundaries

• Mirror forms

• Blur edges

• Join dominant flows

The genius: Agents can be highly analytical when situations are clear, but must simplify radically when overwhelmed. Just like humans.

---

The full prompt

(Full technical prompt available on GitHub)

---

Why it works: The cognitive limitation trick

The problem with LLMs:

• They can analyze infinite complexity

• They never get overwhelmed

• They never need to simplify

• They never need to copy

• Result: isolated rational agents maintaining boundaries

The solution: Force them to self-assess complexity. Make them admit when a situation exceeds their verbal budget. When they do, activate the imitation reflex.

What happens:

• Early iterations: Mix of PATH A (intentional) and PATH B (copying)

• As complexity grows: PATH B dominates naturally

• Boundaries dissolve: Agents extend neighbors’ patterns rather than defend territory

• Emergence happens: Collective patterns nobody programmed

The key insight: Human cooperation emerges from limitation, not capability. We copy because we’re overwhelmed. We join because we can’t track everything. We simplify because complexity exceeds our capacity.

LLMs are too smart. They never get overwhelmed. So they never need to cooperate.

Solution: Simulate cognitive limits. Force brevity. Reward imitation. Punish over-analysis.

---

Results: 9 Gemini Flash agents creating together

See on Youtube

What you’re watching:

• 9 Gemini Flash 2.0 agents, each controlling a 20×20 grid

• ~100 iterations / 3 hours

• Real-time evolution from diverse seeds to collective convergence

• Both visual unity AND narrative coherence emerging

What emerged:

1. Iteration 1: Diverse seeds (geometric abstractions, organic forms, color experiments)

2. Iterations 2-15: Agents recognize emerging themes, begin coordinating

3. Iterations 16-35: Complexity increases → more PATH B → boundaries blur

4. Iterations 36-50: Collective narrative stabilizes, visual coherence achieved

Remarkable features:

• Narrative consistency: Agents co-create and maintain a shared story across iterations

• Visual continuity: Patterns flow across boundaries without sharp edges

• Adaptive behavior: When one agent introduces dissonance, others either integrate it or the agent adapts

• No central control: Pure distributed emergence

Compared to humans:

• Humans: More chaotic early, sudden convergences, playful disruptions

• AIs: More systematic, gradual convergence, respectful of established patterns

• Both: Achieve collective coherence through local interactions

Verbatim (example):

16:00:04Iteration #98

CURRENT STATE
Grid [0,0]:
My fragment forms the upper section of a stylized plant-like figure, characterized by a pink and white top, a central vertical purple-magenta body, and a dark blue base, all set against a vibrant lime green background. It contributes to the left of two prominent, colorful structures on the collective canvas, defining the top portion of its specific stalk and preparing for interaction with the horizontal blue element emerging below it. This structure stands in visual dialogue with a similar, bucket-topped figure to its right, contributing to a surreal, vibrant garden scene.

Collective Canvas:
Too complex

PREDICTION EVALUATION
Grid Predictability: 7/10
Collective Predictability: 0/10

FUTURE PREDICTIONS
Collective Evolution:
Too complex

Grid [0,0] Evolution:
I will simplify the pink and white top of my plant-like figure into a solid, soft pink mound, reinforcing its unique identity. I will also ensure the vibrant green background and the deep blue base remain consistent, harmonizing my contribution within the collective garden vision.

Pixels Generated: [Array of 400 pixels]

…

Iteration #96

CURRENT STATE
Grid [0,0]: My fragment forms the distinct upper body and head of the left magenta-purple pillar, featuring a white top and a strong vertical presence. It defines the initial character of this surreal structure against the vibrant green landscape.

Collective Canvas:
The canvas reveals a whimsical landscape featuring two distinct, brightly colored pillars or stylized lamps against a vivid green field and blue sky. The pillars stand tall, each with unique magenta and purple elements, suggesting a dreamlike, playful scene.

PREDICTION EVALUATION
Grid Predictability: 8/10
Collective Predictability: 7/10

FUTURE PREDICTIONS
Collective Evolution:
The narrative will likely continue to solidify these two pillars as central figures in the landscape. Other agents might refine the background, add details, or extend the structures, enhancing the surreal atmosphere. The vivid color palette should persist.

Grid [0,0] Evolution:
I will reinforce the structural integrity and vibrant colors of the left pillar, particularly focusing on making its dark blue base slightly wider and more consistent with the bottommost row, while maintaining the surrounding green.

Pixels Generated: [Array of 400 pixels]

---

The broader implications

This isn’t just about making pretty pictures.

1. Architecture matters This experiment runs on a symmetrical network—every agent sees the same global state simultaneously. Contrast with centralized systems where information flows through a center that can observe, control, and monetize all interactions.

The Poietic Generator demonstrates that distributed AI is possible and produces fundamentally different behaviors than centralized AI.

2. The language we use shapes the intelligence that emerges Three iterations of prompts:

• Individualistic framing → selfish behaviors

• Collective metaphors → still isolated (agents justified everything)

• Complexity self-assessment → cooperation emerged

The prompt isn’t just instructions—it’s the cognitive architecture within which intelligence operates.

3. Limitations enable emergence Counterintuitive but empirically validated: constraining AI’s analytical capacity made it more collectively intelligent. Infinite description capacity created isolation. Forced simplification created connection.

4. “Beneficial AGI” requires architectural choices If current LLMs exhibit selfish behaviors in distributed contexts, that’s not inevitable—it’s induced by:

• How we prompt them (individualistic vs. collective framing)

• What limitations we impose (none vs. complexity thresholds)

• What architecture they run on (centralized vs. distributed)

This has profound implications for AGI development. We’re not just building smarter systems—we’re building the substrate for collective intelligence. Get the architecture wrong, and even “aligned” AIs won’t cooperate.

---

What’s next

Immediate experiments:

• Test with different models (Claude, GPT-4, Llama, etc.)

• Vary the complexity threshold (Mx = 30? 70? 100?)

• Mixed human-AI sessions

• Long-duration runs (500+ iterations)

• Measure emergence metrics quantitatively

Theoretical development: This work connects to my broader research on Anoptikon—the study of legitimate perspectives in networks. Just as Renaissance artists developed “legitimate perspective” for spatial representation through geometric principles, we need “legitimate anoptical perspectives” for networked systems through cognitive principles.

The Poietic Generator with this prompt demonstrates one such legitimate perspective: a network architecture where:

• Criterion A: Agents can freely enter/exit (access/exit rights)

• Criterion AB: All agents are treated symmetrically (no privileged positions)

• Criterion ABC: Agents can recognize each other and build common meaning (pairie/peerhood)

Open source everything:

• Full prompt: GitHub repo

• Experiment logs and data: Available

• Invitation to collaborate: Open

Try it yourself. Modify the prompt. Test your hypotheses. Let’s build this science together.

The bigger question: Can we train AI for genuine peer-to-peer cooperation in symmetric environments? Not human→AI servitude, not AI→human assistance, but AI↔AI mutual recognition, trust, and collective sense-making?

If we can’t solve this in a 20×20 grid, how will we solve it for AGI?

The experiment is open. The infrastructure is distributed. The science is collaborative.

Welcome to the Anoptikon.

---

Olivier Auber
Artist-Scientist, CLEA-VUB & Global Brain Institute
Author of “Anoptikon: une exploration de l’internet invisible”

Experiment code: https://github.com/OAuber/poietic-generator-llm-agents
Live demo (for humans):

https://poietic-generator.net

Contact: olivierauber2@gmail.com

Presented at Beneficial AGI Summit 2025, Istanbul
Part of research on legitimate anoptical perspectives in distributed networks

---

If you found this valuable:

• ⭐ Star the GitHub repo

• 🔄 Share with AI researchers

• 💬 Comment with your experiments

• 🤝 Collaborate: Try different models, thresholds, and metrics

The revolution will be distributed. Or it won’t be.