By Dr. Volkan Ustun, Director, Human-inspired Adaptive Teaming Systems (HATS) Lab, USC Institute for Creative Technologies
The evolution of artificial intelligence has reached a critical juncture. While the raw computational power of large language models continues to impress, their application to truly complex, multi-faceted problems reveals fundamental limitations that single-agent approaches cannot overcome. At the USC Institute for Creative Technologies, our Human-inspired Adaptive Teaming Systems (HATS) Lab has been pioneering a different path forward—one that mirrors the collaborative intelligence found in human teams rather than relying on monolithic AI systems.
Our recent work demonstrates that the future of AI lies not in building ever-larger singular models, but in orchestrating specialized agents that can reason, adapt, and coordinate toward common objectives. This shift from individual to collective intelligence represents more than a technical advancement; it embodies a fundamental reimagining of how artificial systems can tackle the kind of complex, real-world challenges that have long been the exclusive domain of human expertise.
It’s important to distinguish our work from what popular media often describes as “agentic AI”—systems that can independently navigate web browsers, make purchases, or interact broadly with external digital environments. Our research focuses on domain-specialized agentic systems that operate within carefully bounded problem spaces: physics simulation environments and structured military planning frameworks. While our agents demonstrate genuine autonomy in reasoning, planning, and action-taking within these domains, they are designed for controlled, high-stakes applications rather than general-purpose interaction with the broader digital world. This constrained approach allows for deeper specialization, more predictable behavior, and enhanced safety—characteristics essential for military training applications where precision and reliability are paramount.
The Limitations of Single-Agent Thinking
Traditional AI approaches, even those powered by the most sophisticated large language models, struggle with what we call the “monolithic reasoning problem.” When faced with complex tasks requiring multiple perspectives, specialized knowledge, and iterative refinement, single-agent systems often fall into predictable patterns of failure. They may generate plausible-sounding solutions that lack internal coherence, struggle to maintain consistency across multiple interdependent components, or fail to adapt their reasoning when confronted with feedback from real-world constraints.
This limitation became particularly apparent in our work on automated scenario generation for military training. The challenge of creating realistic, complex training scenarios requires simultaneous consideration of geographic constraints, force dynamics, tactical objectives, and doctrinal requirements. No single AI agent, regardless of its underlying model’s sophistication, can effectively manage all these dimensions while maintaining the kind of specialized expertise needed in each domain.
Consider the task of generating a military operations order—a document that must seamlessly integrate unit movements, terrain analysis, logistical considerations, and contingency planning. Each component requires distinct types of reasoning: spatial analysis for movement planning, temporal reasoning for coordination, tactical knowledge for force employment, and narrative coherence for clear communication. Attempting to handle all these requirements within a single prompting framework inevitably leads to compromises in quality and consistency.
A New Paradigm: Distributed Intelligence
Our multi-agent approach fundamentally reconceptualizes how AI systems can approach complex problems. Rather than asking a single agent to be expert in everything, we create specialized agents that excel in particular domains and coordinate their efforts through structured interaction protocols. This mirrors how human teams naturally organize around areas of expertise while maintaining coherent communication channels.
In our physics puzzle generation research, we developed a framework that employs multiple ReAct agents—systems that integrate reasoning and action in iterative loops. The architecture includes a central coordinator that orchestrates the overall process, while specialized agents handle specific aspects of the problem. A Designer agent focuses on spatial planning and object placement, while a Solver agent concentrates on ensuring that generated puzzles remain solvable. Each agent brings focused expertise to bear on its particular domain while contributing to a larger, coherent solution.
The results speak to the power of this approach. Where single-agent systems struggled with spatial reasoning errors and inconsistent puzzle generation, our multi-agent framework demonstrated significantly improved performance in creating solvable puzzles that aligned with user specifications. More importantly, the system exhibited the kind of adaptive reasoning that allows for real-time adjustment based on feedback from simulated environments.
From Physics to Military Strategy
The principles we validated in the physics puzzle domain translate directly to much more consequential applications. Military training scenarios represent one of the most demanding tests of AI-assisted content generation, requiring not only technical accuracy but also strategic coherence, operational realism, and adherence to complex doctrinal frameworks.
Working in collaboration with experts from the Army University, we have developed a multi-agent framework specifically designed to automate the generation of military training artifacts, including operations orders (OPORDs). The system decomposes the complex task of scenario generation into a hierarchy of interdependent subproblems, each handled by specialized agents with domain-specific expertise.
Our framework includes agents responsible for different aspects of military planning: analyzing learning objectives, generating force compositions, reasoning about terrain and movement, developing backstories and political contexts, and synthesizing these elements into coherent doctrinal documents. The system can process both textual inputs and geospatial data, enabling it to reason about unit positions, movement paths, and tactical objectives in realistic operational contexts.
The technical architecture centers on an Orchestrator Agent that manages the overall generation process while delegating specialized tasks to Helper Agents. Each helper agent focuses on a particular information block—whether that’s map analysis and terrain reasoning, unit positioning over time, or decision support matrices. The orchestrator integrates outputs from these specialized agents to produce coherent, context-aware documents that maintain logical consistency across all components.
Overcoming Spatial and Temporal Reasoning Challenges
One of the most significant technical challenges we’ve addressed is the integration of spatial and temporal reasoning capabilities. Traditional language models, trained primarily on textual data, often struggle with precise coordinate systems and dynamic movement prediction. Our multi-agent approach addresses this limitation through specialized agents equipped with discretized movement systems and graph-based reasoning capabilities.
The Map and MCOO (Modified Combined Obstacle Overlay) agent, for instance, doesn’t just provide static terrain information but serves as an interactive visual reasoning tool. It can generate waypoint-based movement systems, identify multiple viable paths between locations, and provide progress-tracking visualization that enables other agents to reason about unit positions over time. This capability proved essential for generating realistic unit movement predictions and coherent schemes of maneuver.
Similarly, our Unit Positions agent combines tactical knowledge with temporal reasoning to predict how military units will move through complex operational environments. Rather than simply calculating shortest paths, the system incorporates contextual factors like terrain vulnerability, operational security considerations, and coordination requirements with other units.
Beyond Technical Achievement: Implications for Human-AI Collaboration
The significance of this work extends beyond its immediate technical achievements. Our multi-agent framework represents a new model for human-AI collaboration—one where artificial systems don’t simply replace human expertise but augment and extend human capabilities in sophisticated ways.
In our military applications, we’ve categorized different types of AI assistance based on the level of human oversight required. Some tasks, like generating multiple options for human selection, require minimal automation but significant human judgment. Others, like producing candidate documents for human approval, benefit from more extensive AI involvement while maintaining human oversight. Still others, particularly routine textual synthesis tasks, can be handled largely automatically while preserving human review capabilities.
This nuanced approach to automation reflects a deeper understanding of where AI systems excel and where human judgment remains irreplaceable. Rather than pursuing full automation as an end goal, we’re developing systems that enhance human decision-making by handling routine cognitive tasks, generating multiple alternatives for consideration, and maintaining consistency across complex document structures.
Persistent Challenges and Future Directions
Despite these advances, significant challenges remain. Current language models still struggle with precise spatial coordinate generation, occasionally exhibit directional confusion in visual analysis, and sometimes prioritize technical solvability over alignment with user intent. These limitations point to fundamental gaps in how language models process spatial information and translate high-level reasoning into precise, actionable outputs.
Our ongoing research addresses these challenges through several approaches. We’re investigating fine-tuning strategies to improve spatial reasoning capabilities, developing evaluative agents that can assess whether generated content meets specific requirements, and exploring how newer model architectures might better handle the kind of grounded reasoning required for real-world applications.
Perhaps most importantly, we’re working to ensure that our multi-agent systems remain transparent and interpretable to human operators. The distributed nature of multi-agent reasoning can make it challenging to understand how particular decisions emerge from the interaction of multiple specialized components. Developing clear visualization and explanation capabilities will be crucial for maintaining appropriate human oversight as these systems become more sophisticated.
The Future of Collaborative Agentic Intelligence
The implications of agentic AI extend far beyond military applications. Any domain requiring the integration of multiple types of expertise, the coordination of complex workflows, or the synthesis of information from diverse sources could benefit from similar approaches. Educational content generation, healthcare protocol development, infrastructure planning, and scientific research coordination all present opportunities for agentic AI systems to augment human capabilities through autonomous reasoning and adaptive action-taking.
The key insight driving our work is that intelligence—whether artificial or natural—is fundamentally collaborative and agentic. Human expertise emerges not from individual brilliance but from the accumulated knowledge of communities, the refinement of ideas through peer review, and the synthesis of diverse perspectives into coherent solutions. However, human intelligence is also inherently agentic—it involves goal formation, planning, autonomous action, and adaptation based on outcomes. Our agentic AI systems attempt to capture both the collaborative and autonomous aspects of intelligence in computational form.
In the future, the most promising AI systems won’t be those that try to replicate human intelligence in singular, reactive forms. Instead, they’ll be those that can autonomously orchestrate specialized capabilities, adapt to complex constraints, and collaborate effectively with both human experts and other agentic AI systems. This vision of distributed, collaborative, and autonomous intelligence offers a path toward AI systems that are not only more capable but also more aligned to constructive human-AI teaming.
The work detailed in our EMAS and I/ITSEC papers represents just the beginning of this transformation. As agentic AI systems become more sophisticated and their applications more diverse, they promise to reshape how we approach complex problem-solving across virtually every domain of human endeavor. The future belongs not to artificial intelligence that replaces human thinking, but to agentic systems that amplify human intelligence through sophisticated collaboration, autonomous reasoning, and specialized expertise working in concert.
//