Source-linked AI summary
Human-Artificial Interaction in the Age of Agentic AI: A System-Theoretical Approach
Uwe M. Borghoff, Paolo Bottoni, Remo Pareschi
TL;DR
HCI needs a systems perspective for coordinating heterogeneous human and computational agents beyond traditional interfaces. The paper distinguishes autonomous multi-agent coordination from Centaurian integration and develops Petri-net-based communication spaces to formalize both. Its framework is positioned for agentic-AI collaboration that preserves autonomy where appropriate while supporting tight human–AI integration and adaptive operation.
Problem
Traditional HCI does not fully capture coordination among heterogeneous human and computational agents with different capabilities, roles, and goals.
Method
The paper develops a Petri-net-based communication-spaces framework for modeling both autonomous multi-agent coordination and tightly integrated Centaurian collaboration.
Results
The framework supports agentic-AI interactions ranging from autonomous coordination to tight human–AI integration, including routine operation and complex human-guided decisions.
Takeaways & Limitations
The framework provides a theoretical foundation for designing hybrid human–AI systems that combine defined coordination protocols with deeper functional integration.
Abstract
from arXiv · showhide
This paper presents a novel perspective on human-computer interaction (HCI), framing it as a dynamic interplay between human and computational agents within a networked system. Going beyond traditional interface-based approaches, we emphasize the importance of coordination and communication among heterogeneous agents with different capabilities, roles, and goals. A key distinction is made between multi-agent systems (MAS) and Centaurian systems, which represent two different paradigms of human-AI collaboration. MAS maintain agent autonomy, with structured protocols enabling cooperation, while Centaurian systems deeply integrate human and AI capabilities, creating unified decision-making entities. To formalize these interactions, we introduce a framework for communication spaces, structured into surface, observation, and computation layers, ensuring seamless integration between MAS and Centaurian architectures, where colored Petri nets effectively represent structured Centaurian systems and high-level reconfigurable networks address the dynamic nature of MAS. Our research has practical applications in autonomous robotics, human-in-the-loop decision making, and AI-driven cognitive architectures, and provides a foundation for next-generation hybrid intelligence systems that balance structured coordination with emergent behavior.
1 Introduction
Agentic AI expands HCI into interactions among heterogeneous human and computational agents, motivating a distinction between autonomous multi-agent coordination and deeper Centaurian integration. The paper proposes Petri-net-based communication spaces to formalize both paradigms.
- Agentic AI systems support iterative planning, autonomous task decomposition, and continuous learning, broadening the scope of human–computer interaction.
- Powerful conversational LLMs enable humans to participate as fully capable agents in complex, interconnected multi-agent ecosystems.
- Multi-agent systems preserve autonomous entities and roles, whereas Centaurian systems fuse human and machine capabilities more deeply.
- The proposed Petri-net framework models heterogeneous agents while supporting clear boundaries, regulated interactions, and adaptive feedback loops.
- The paper progresses from paradigm analysis and Petri-net foundations to communication spaces and practical HCI use cases.
2 Paradigms in Human-AI Integration
The paper distinguishes MAS, which coordinate autonomous agents through defined protocols, from Centaurian systems, which functionally integrate human and artificial components. It then situates both within agentic-AI collaboration, including adaptive settings where the modes can complement one another.
- Multi-agent systems: MAS preserve distinct agent boundaries and identities while enabling complex interactions through autonomous decision-making and coordination.
- Centaurian systems: Centaurian systems form unified composite entities in which human and artificial components become functionally interdependent.
- Communication spaces: Communication spaces support information exchange, coordination, common ground, inter-predictability, and directability across both paradigms.
- Operational differences: MAS adapt through relationship reconfiguration, whereas Centaurian systems evolve through internal transformation and permeable integration boundaries.
- Convergence: Large action model interactions can appear as MAS collaboration or Centaurian integration depending on context, allowing the paradigms to complement each other.
- System design: Communication spaces are designed to support MAS loose coupling and Centaurian tight integration while preserving each paradigm’s core attributes.
- Agentic AI collaboration: Agentic AI can operate autonomously for routine tasks and integrate tightly with human experts for complex decisions within the same enterprise system.
- Agentic AI collaboration: Continuous learning and adaptation apply both to improving independent agents and to refining human–AI integration patterns.
3 Technical Background: Petri nets
Petri nets provide a graphical and mathematically precise formalism for concurrent processes, synchronization, and token-based state changes. Colored extensions add typed tokens and guards, making the framework suitable for heterogeneous human–AI interactions and communication spaces.
- Petri-net foundations: Their graphical representation, explicit concurrency, formal analysis tools, and modular extensions support rigorous modeling of distributed and human–AI workflows.
- Petri-net foundations: Petri nets model concurrent-system information and control flow through places, transitions, arcs, and tokens that move as transitions fire.
- Colored Petri nets: Colored Petri nets assign types or colors to tokens, turning places into typed containers and enabling guarded transitions with more expressive behavior.
- Colored Petri nets: Colored tokens can represent heterogeneous messages, tasks, and capabilities associated with human and synthetic agents.
- Colored Petri nets: Color sets track agent states and contexts, allowing transitions to fire only when specified guard conditions are satisfied.
- Communication-space modeling: Colored Petri nets specify interaction protocols, including dynamic task creation and distribution, and map communication spaces onto distinct typed domains.
- Illustrative examples: A simple Petri net depicts places as circles, transitions as rectangles, and arcs as token-flow connections in a concurrent process.
- Illustrative examples: A colored Petri net uses typed tokens or data to represent agent roles or tasks and enable conditional transitions.
4 Communication Spaces and Agent Architectures
Communication spaces provide a unified framework for coordinating heterogeneous agents across multi-agent and Centaurian architectures. The framework organizes interactions into layered spaces and extends Petri nets to represent typed communication, synchronization, feedback, and shared decision-making.
- Communication Spaces: Communication spaces address the need to support both loose coupling in multi-agent systems and deep integration in Centaurian systems.The framework extends Petri-net-based modeling to capture heterogeneous capabilities, communication patterns, and changing coordination requirements.
- Heterogeneous Communication: A human operator, LLM agent, and drone swarm can exchange text, structured sensor data, and video through channels with different synchronization, data-format, and reliability requirements.Communication spaces formalize these heterogeneous channels while preserving the distinctions among strategic decisions, real-time state updates, and observation.
- Layered Architecture: The architecture organizes interactions into surface, observation, and computation spaces, each with distinct interaction rules and constraints.Surface space mediates interfaces, sensors, and APIs; observation space transforms and routes messages; computation space performs decision-making and resource allocation.
- Layered Architecture: The layered architecture links surface agents, observer agents, and executive agents, with observation agents formatting data for presentation and decoding user interactions.This organization resembles the Model-View-Controller pattern and connects external interaction with internal computation.
- Group Agents: Group-agents manage communication spaces by coordinating users and agents, tracking active membership, and supporting distributed problem-solving.Their structure includes shared identity and status information, group-specific composition, and behaviors for message delivery and membership changes.
- Formal Representation: Communication Space Petri nets partition places into three layers, use typed tokens for data formats or message semantics, and apply capability-aware transition guards.The unified model gives designers a blueprint for maintaining agent boundaries while permitting deeper integration and emergent capabilities when appropriate.
5 Use Case 1: Multi-agent Interaction with Satellite and Swarm Robots
The satellite-and-swarm use case combines autonomous agents with tightly integrated human–LLM decision making. Its semi-centralized architecture supports coordinated, adaptive operation while preserving independent agent roles and human oversight.
- System architecture: Information and decisions flow through a semi-centralized architecture that integrates human operators, conversational AI, satellite control, and swarm robots.Figure 8 presents the interaction, while Figure 9 presents the overall information and decision-making flow.
- System architecture: The system combines a satellite control unit, swarm robots, an LLM, and human operators as distinct autonomous components.Each component retains decision-making capabilities while participating in a collective process.
- Hybrid collaboration: The integration layer combines multi-agent independence with Centaurian human–LLM coupling, forming a tightly integrated decision-making unit.The LLM remains an independent intelligence layer when interacting with control algorithms, while human–LLM interaction integrates judgment and recommendations.
- Practical applications: In path optimization and adaptive problem solving, the LLM advises independently while human review temporarily shifts interaction into a Centaurian mode.For path planning, the LLM can suggest curve-parameter changes in response to obstacles or terrain, and operators review those suggestions.
- Practical applications: Supporting both paradigms produces a robust system that combines coordinated independence with deep human–AI integration and protects against automation risks.The collaboration model is designed to leverage the distinct strengths of multi-agent and Centaurian approaches.
6 Use Case 2: Large Action Models (LAMs) through Feedback Loops on HCI Interactions
The LAM use case illustrates a Centaurian HCI model in which human and computational components evolve together through feedback. Its distributed architecture retains multi-agent elements while enabling predictive, adaptive, and proactive interaction.
- System paradigm: Rabbit Tech’s LAM combines a Centaurian emphasis on tight human–AI integration with multi-agent elements in a distributed architecture.The framework models and predicts human actions on computer applications.
- Feedback-loop integration: LAM continuously learns from user input and refines predictive capabilities, making boundaries between human and artificial intelligence increasingly fluid.This feedback-loop design represents a dynamic form of Centaurian integration rather than fixed component boundaries.
- Computational integration: The LAM framework forms a unified decision-making entity through neuro-symbolic programming that integrates computational processing with human users and the HCI system.Its core components include the LAM node, human users, and the HCI system.
- Feedback-loop integration: Natural-language interaction and feedback allow LAM to adapt its models to user preferences as human intent and machine execution become increasingly blurred.The interface accepts spoken or typed commands and updates through each interaction.
- Practical applications: LAM can anticipate user needs and provide proactive solutions, but tight integration creates challenges for decision transparency and privacy.These challenges arise where human and artificial boundaries become less distinct.
- Key takeaways: Feedback loops let human and artificial components grow together, creating more intuitive and responsive computing environments while retaining some distributed multi-agent characteristics.The system continuously refines interaction patterns rather than relying solely on stable protocols.
7 Related Work
Related work spans formal models of agent coordination, human–AI interaction studies, distributed architectures, and ethical critiques. Together, these strands position the paper within research on interaction protocols, adaptive systems, and the consequences of AI collaboration.
- Agent interaction models: Prior work formalizes distributed computation, knowledge combination, distributed problem solving, and intention recognition among agents.The literature has progressively addressed higher-level interactions as artificial agents become more sophisticated.
- Distributed architectures: Related architectural work addresses mobile applications, satellite and swarm robotics, IoT systems, and distributed programming across network tiers.Examples include automated architectural analysis, semi-centralized swarm coordination, and distributed MVC extensions for IoT.
- Human–AI interaction studies: HCI research reports both mutual benefits from AI collaboration and negative experiences, including higher frustration without productivity or self-efficacy gains in one ChatGPT study.The cited study linked frustration to errors violating human–AI interaction guidelines.
- Human–AI interaction studies: Other studies examine generative AI in software engineering, task allocation between developers and AI, user reactions, and anthropomorphic responses to chatbots.These works investigate where AI may assist people and how users interpret artificial agents.
- Computational approaches: Message Passing Neural Networks provide a related computational model in which graph nodes learn representations by aggregating information from neighboring nodes.The paper connects this message-passing paradigm to the Rabbit use case.
- Agent interaction models: Concurrent game structures and temporal logics model agents selecting actions from global system states, while other approaches incorporate spatial constraints.These formalisms address possible system evolutions under individual agent strategies.
- Ethical perspectives: The literature also includes humanistic and ethical perspectives concerning human achievement, human–AI relations, and possible existential risks.These perspectives are identified as beyond the paper’s scope.
8 Conclusion and Future Work
The paper concludes that HCI can be analyzed as a networked interplay among human and artificial agents, supported by two use cases. Future work targets more adaptable MAS and a hybrid formal framework connecting dynamic autonomy with structured integration.
- Conclusion: The proposed framework treats HCI as dynamic interaction among multiple agents within a networked ecosystem and applies it to satellite–swarm robotics and LAMs.The two use cases support the framework’s practical implications.
- Use-case findings: In the satellite–swarm use case, an LLM-enabled semi-centralized system improves decision making and adaptability through real-time data-driven adjustments while preserving human oversight.Human oversight is described as critical for complex, dynamic environments.
- Use-case findings: In the LAM use case, neuro-symbolic programming enables operations to refine continuously from human feedback, producing an adaptive and responsive HCI system.The result concerns prediction and shaping of human–computer interactions.
- Conclusion: The framework is intended to improve understanding of collective agent capabilities, system adaptability, and resilience by emphasizing complex human–machine interplay.The conclusion presents this as a foundational approach for HCI design and analysis.
- Future work: Future MAS research will address dynamic team composition, because static Petri-net structures have difficulty accommodating agents joining or leaving while maintaining consensus.The proposed direction models agents as individual colored Petri nets with specialized input and output locations.
- Future work: A hybrid framework will use communication spaces to connect structured Centaurian intelligence, modeled by Petri nets, with decentralized MAS interactions supported by high-level reconfigurable networks.The goal is to enable dynamic transitions between stable coordination structures and self-organizing agent interactions.
- Future work: Set-theoretic models are proposed as a complementary route for abstracting dynamic agent relationships while retaining structured protocols for message flow and consistency.This direction is intended to complement Petri nets rather than replace formal structure altogether.
Funding
The paper acknowledges funding for Remo Pareschi from the European Union—NextGenerationEU through an Italian Ministry of University and Research grant.
- Remo Pareschi received funding from the European Union—NextGenerationEU.
- The funding was provided under the Italian Ministry of University and Research National Innovation Ecosystem grant ECS00000041-VITALITY.
- The acknowledged grant carries the identifier CUP E13C22001060006.