Building Socio-Affective Artificial Intelligence for Interactive Multi-Agent Simulations

A software architecture for socially-aware, emotion-driven agents in shared simulated worlds

David Berga· 2026· DOI 10.48550/arXiv.2609.26927

The core problem

The article addresses the challenge of designing interaction between humans and multiple artificial agents within simulated dynamic worlds. It situates this problem at the intersection of general artificial intelligence (AI/AGI), the proliferation of transformer-based conversational agents, and increasing computational capabilities. The central premise is that an integrative design of agent interactions—both among agents themselves and with humans—is crucial for understanding how to create sustainable governance in future human-agent reasoning systems. The work draws on current and previous multi-agent theories of mind, specifically socially and affectively-aware agents, to argue that socio-affective reasoning must be embedded in agent behavior and interaction. The proposed solution is a software architecture called AGIMUD, which enables dynamic world recreation as multi-user dungeons (MUDs) where both artificial agents and humans can interact simultaneously in real time.

The primary result is the AGIMUD software architecture itself, which successfully integrates the three specified components: socially-aware reasoning and emotion, multimodal human-agent interaction, and distributed AI processing. The architecture enables the recreation of dynamic worlds as multi-user dungeons where multiple autonomous agents and human users interact simultaneously in real time. The open-source implementation demonstrates the feasibility of the design principles. While the abstract does not report quantitative performance metrics, the qualitative outcome is a working system that supports the stated integrative goals. The code repository at https://github.com/dberga/AGIMUD serves as a concrete artifact for reproducibility and further development. The results suggest that embedding socio-affective capabilities in agents within a distributed MUD framework is a viable approach for studying human-agent reasoning systems.

The article addresses the challenge of designing interaction between humans and multiple artificial agents within simulated dynamic worlds. It situates this problem at the intersection of general artificial intelligence (AI/AGI), the proliferation of transformer-based conversational agents, and increasing computational capabilities. The central premise is that an integrative design of agent interactions—both among agents themselves and with humans—is crucial for understanding how to create sustainable governance in future human-agent reasoning systems. The work draws on current and previous multi-agent theories of mind, specifically socially and affectively-aware agents, to argue that socio-affective reasoning must be embedded in agent behavior and interaction. The proposed solution is a software architecture called AGIMUD, which enables dynamic world recreation as multi-user dungeons (MUDs) where both artificial agents and humans can interact simultaneously in real time.

The methodology centers on the design and integration of three core components within the AGIMUD software architecture:

The discussion emphasizes that the existence of an integrative design of agent interactions—both self-interactions and interactions with humans—is crucial for understanding how to create sustainable and governable future human-agent reasoning systems. By combining socially-aware reasoning, emotion, multimodal interfaces, and distributed processing, AGIMUD addresses the complexity of real-time multi-agent simulations. The architecture connects the current era of general AI/AGI with the proliferation of transformer-based conversational agents and increased computational capabilities. The use of MUDs as a testbed allows for dynamic world recreation where both agents and humans can interact simultaneously, providing a platform for observing emergent social and affective behaviors. The work implies that sustainable governance in such systems requires not only technical integration but also careful design of interaction protocols and affective models. Future research could extend the architecture to larger-scale networks, incorporate more sophisticated theory of mind models, and evaluate the system with quantitative user studies. The open-source release invites community contributions to refine and expand the design principles.

CS practitioners and researchers

Opening member content…