Constructive Adversarial Architecture: Overcoming Cooperation Bias in Autonomous Multi-Agent Narrative Systems

• Publication: Zenodo (open access preprint)
• Author: Max Capacity
• Date: March 30, 2026
• Type: Author/Publication
• DOI: 10.5281/zenodo.19325616
• PDF: 47 pages, 8.5 MB

Summary

Research paper authored by Max Capacity documenting a system where nine AI agents play Dungeons & Dragons autonomously: a DM, three player characters, a rules enforcer, and post-session agents that write narratives, build a wiki, and publish to a live website. The system runs unattended and produces coherent 20-session campaigns for about $17 USD each on DeepSeek.

The paper identifies cooperation bias — the DM resolves every hostile encounter through diplomacy regardless of instructions — and documents six categories of fixes tested across nine controlled campaigns (200+ sessions, ~$155 USD total). The central finding is that guard rails (instructions that prohibit behavior) degrade over time, while guide rails (structural constraints that produce behavior) work because the unwanted outcome becomes impossible. Boss fight rates went from 25% to 100% using an adversarial architecture approach.

The finding generalizes: in any multi-agent system where one AI controls other entities, give every participant an independent voice rather than constraining the controller.

Significance

This is Max Capacity’s first academic research publication. It sits at the intersection of his art practice (AI, systems, generative processes) and AI research. Published as an open-access preprint on Zenodo with a DOI, indexed in OpenAIRE.

Works Mentioned

(No specific art works referenced — this is original AI research.)

Links

• Zenodo record
• DOI: 10.5281/zenodo.19325616

---

← PRESS