Abstract
Teamwork in complex, dynamic, multi-agent domains mandates highly flexible coordination and communication. Simply fitting individual agents with precomputed coordination plans will not do, for their inflexibility can cause severe failures in teamwork, and their domain-specificity hinders reusability. Our central hypothesis is that the key to such flexibility and reusability is agent architectures with integrated teamwork capabilities. This fundamental shift in agent architectures is illustrated via an implemented candidate: STEAM. While STEAM is founded on the joint intentions theory, practical operationalization has required it to integrate several key novel concepts: (i) team synchronization to establish joint intentions; (ii) constructs for monitoring joint intentions and repair; and (iii) decision-theoretic communication selectivity. Applications in three different complex domains, with empirical results, are presented.
Introduction
In the late 1990s, multi-agent systems—particularly those in high-stakes environments like military simulations or competitive sports—relied heavily on "precomputed coordination plans." Developers essentially hard-coded every possible interaction, a practice that led to two significant issues. First, these systems were inherently inflexible; if a teammate failed or environmental conditions shifted, agents often "hung" because they lacked the capacity to reason about the team's collective state. Second, the coordination logic was highly domain-specific, making it nearly impossible to reuse code across different applications.
True teamwork requires more than just individual flexibility; it necessitates that agents explicitly represent and reason about team goals, mutual beliefs, and joint commitments. To address these challenges, we introduce STEAM (a Shell for TEAMwork), an agent architecture designed to move beyond brittle scripts toward a general, flexible model of teamwork.
The STEAM Architecture
STEAM is built upon the foundations of Joint Intentions Theory, but extends it for practical, real-time application. The architecture introduces several core technical innovations, most notably the concept of hierarchical team operators.
Team Operators and Synchronization
Unlike standard operators that represent individual actions, STEAM utilizes hierarchical "team operators." When a team operator is selected, the architecture automatically triggers a synchronization protocol—a "request-confirm" logic—to establish Joint Persistent Goals (JPGs). This ensures that all team members are committed to the joint goal before execution begins.
Monitoring, Repair, and Communication
STEAM employs constraints (such as AND, OR, and specific dependencies) to monitor team performance. If a critical failure is detected, such as a lead scout being incapacitated, the architecture can automatically trigger a "repair" operator to reconfigure the team.
To manage communication overhead, STEAM incorporates decision-theoretic selectivity. Agents use a decision tree to calculate the Expected Utility (EU) of communicating a piece of information. A message is only sent if the risk of miscoordination outweighs the communication cost, effectively filtering out redundant information that is likely already common knowledge.
Evaluation Domains
The effectiveness of STEAM was evaluated across three distinct and complex simulation environments: the Attack domain, the Transport domain, and RoboCup synthetic soccer. The Attack domain, in particular, provided a high-fidelity testbed for complex battlefield maneuvers.
Key Insights and Results
The shift to treating teamwork as an architectural primitive yielded significant improvements in both flexibility and reusability:
- Enhanced Flexibility: STEAM-based agents successfully avoided the "hanging" failures characteristic of earlier systems. They could autonomously detect when a mission became unachievable and coordinate a retreat or reconfiguration.
- Communication Efficiency: In the Attack domain, agents using decision-theoretic selectivity performed as well as those that communicated everything, but used 10 to 20 times fewer messages. This efficiency was critical for maintaining real-time performance.
- High Code Reuse: Researchers achieved 100% reuse of STEAM's coordination rules when transitioning from the Attack domain to the Transport domain. Even in the vastly different RoboCup environment, 40% of the rules were reused without modification.
Conclusion and Future Work
While STEAM demonstrates the power of integrated teamwork capabilities, challenges remain. In the RoboCup domain, lower reuse rates were attributed to difficulties in spatial reasoning, which impacted role-monitoring logic. Future work will focus on learning decision-theoretic parameters over time and utilizing "chunking" (explanation-based learning) to transition deliberate teamwork reasoning into automatic, situation-specific rules. By embedding teamwork into the agent's core architecture, STEAM provides a path toward more robust and reusable multi-agent systems.
Summary updated