Human-Artificial Interaction in the Age of Agentic AI
Summary: Borghoff, Bottoni & Pareschi (2025) apply Luhmann’s systems theory and von Bertalanffy’s living-systems framework to human-AI interaction, drawing a sharp distinction between Multi-Agent Systems (ecosystem model) and Centaurian systems (organism model), and formalizing both with colored Petri nets.
Sources: Academia/2025FRONTIERSHuman-ArtificialInteraction.pdf
Last updated: 2026-05-06
The two paradigms
Agentic AI has revived two fundamentally different models of human-machine collaboration:
Multi-Agent Systems (MAS) — the ecosystem model
Agents — human or artificial — maintain distinct boundaries and autonomous decision-making while coordinating via structured protocols. Like biological ecosystems: organisms interact without merging. Key properties:
- Identity maintenance: each agent remains a distinct entity
- Adaptation: via reconfiguration of agent relationships
- Boundaries: clear, well-defined interfaces
Centaurian systems — the organism model
Human and artificial components merge into a functionally unified entity. The term “Centaurian” (after Pareschi 2024) evokes the mythological fusion — and moves beyond Licklider’s original vision of human-computer symbiosis. Properties:
- Identity: a new composite identity is created through integration
- Adaptation: through internal transformation, not external reconfiguration
- Boundaries: permeable, blurring human decision-making and AI processing
The distinction echoes a parallel in biology: ecosystems vs. organisms. Both involve interacting intelligent entities; they represent fundamentally different structural logics. (source: Academia/2025FRONTIERSHuman-ArtificialInteraction.pdf)
From Homo Faber to Centaurus Faber
Herbert Simon’s tripartite cognitive architecture — external interface, coding mechanism, internal processing — was conceived for human problem-solving but extends naturally to hybrid systems. The trajectory:
- Homo Faber: human uses AI as a tool (traditional HCI)
- Centaurus Faber: human and AI merge into a new kind of cognitive agent
The Centaurian configuration is not merely technological fusion but a new ontological category — a system that can be instantiated by either human cognition or AI subsystems depending on context and task.
Communication spaces
Both paradigms share the concept of communication spaces — regions of interaction where information exchange and coordination occur. The paper structures these into three layers:
- Surface layer — the visible interface between agents (what humans see and interact with)
- Observation layer — monitoring, sensing, and representing the state of the system
- Computation layer — internal processing, reasoning, and decision-making
These layers enable what the authors call “joint activity”: characterized by inter-predictability, common ground, and directability.
Formal modeling: colored Petri nets
The paper formalizes both paradigms using colored Petri nets — a modeling language that supports parallel processes and state transitions via token semantics. Unlike graph-based frameworks, Petri nets natively handle dynamic concurrency, critical for hybrid systems where multiple agents (human and artificial) act simultaneously.
The formalism supports:
- Encoding heterogeneous data and agent roles
- Modeling learning, adaptation, and reasoning
- Auditability — traceable paths of action for ethical accountability
The central tension
MAS preserves autonomy; Centaurian systems sacrifice it for integration. Neither is categorically superior — context determines which is appropriate:
- MAS: better when agent independence matters (e.g., distributed robotics, multi-party decision-making)
- Centaurian: better when seamless fusion enables capabilities neither party could achieve alone (e.g., cognitive prosthetics, co-creative AI)
Large action models (LAMs) — where humans provide training feedback — sit at the convergence: depending on context, the same interaction can manifest as MAS collaboration or Centaurian integration.
Ethical implications
The blurring of boundaries between human and artificial agents raises accountability questions. When decision-making authority is shared in healthcare, defense, or finance, traceable paths of action become essential. The paper’s formal architecture supports auditability even in deeply integrated Centaurian configurations.
Connections
systems-thinking — Luhmann and von Bertalanffy both inform the living-systems framework here. The MAS/Centaurian distinction maps onto Meadows’s structural analysis: changing interconnections (MAS) vs. changing purpose/identity (Centaurian) as different levels of systemic intervention.
technology-and-humans — Pirsig argues technology is a fusion of nature and the human spirit, not an exploitation. The Centaurian model takes this literally: the human-AI hybrid is a new kind of creation that transcends both.
llm-and-mind — the Centaurian model raises questions the other papers in that synthesis don’t address. If LLMs lack original intentionality (Borg) and relevance realization (Jaeger), can a Centaurian system built from them plus a human acquire these properties at the system level? The paper doesn’t resolve this, but it sharpens the question.