This article presents a general model of teamwork called STEAM, designed to enable flexible and coherent teamwork in complex, dynamic multi-agent environments. The central hypothesis is that providing agents with general models of teamwork allows them to autonomously reason about coordination and communication, providing the flexibility needed to address uncertainties in such environments. STEAM is based on joint intentions theory, which provides a principled framework for reasoning about communication and coordination. It also parallels and borrows from the SharedPlans theory, which emphasizes the importance of shared goals and plans in teamwork. STEAM's key features include the use of joint intentions as the basic building block of teamwork, the ability to monitor team performance and reorganize as needed, and decision-theoretic communication selectivity to reduce communication overheads. The model is implemented in a way that allows agents to reason about their coordination and communication responsibilities, and to anticipate and recover from teamwork failures. STEAM has been tested in three different complex domains: Attack, Transport, and RoboCup synthetic soccer. In the Attack domain, STEAM was used to manage a team of synthetic attack helicopters, where agents had to coordinate and communicate to achieve a common goal. In the Transport domain, STEAM was used to manage a team of synthetic transport helicopters, where agents had to coordinate and communicate to ensure the safe transport of troops. In the RoboCup synthetic soccer domain, STEAM was used to manage a team of synthetic soccer agents, where agents had to coordinate and communicate to achieve a common goal. The results of these experiments showed that STEAM was effective in enabling flexible and coherent teamwork in these complex, dynamic environments. The model was able to handle uncertainties, such as differing views of the environment, unexpected failures, and unexpected opportunities. STEAM's ability to monitor team performance and reorganize as needed was particularly important in these environments. STEAM's implementation is based on the Soar architecture, with a set of about 300 domain-independent Soar rules. Three different teams have been developed based on this operationalization of STEAM. These teams have a complex structure of team-subteam hierarchies and operate in complex environments. The article presents detailed experimental results from these teams, illustrating the benefits of STEAM in their development. STEAM is among just a very few implemented general models of teamwork. Other models include Jennings' joint responsibility framework in the GRATE system and Rich and Sidner's COLLAGEN, which both operate in complex domains. While Section 7 will discuss these in greater detail, STEAM significantly differs from both these frameworks, via its focus on a different (and arguably wider) set of teamwork capabilities that arise in domains with teams of more than two-three agents, with more complex team organizational hierarchies, and with practical emphasis on communication costs.This article presents a general model of teamwork called STEAM, designed to enable flexible and coherent teamwork in complex, dynamic multi-agent environments. The central hypothesis is that providing agents with general models of teamwork allows them to autonomously reason about coordination and communication, providing the flexibility needed to address uncertainties in such environments. STEAM is based on joint intentions theory, which provides a principled framework for reasoning about communication and coordination. It also parallels and borrows from the SharedPlans theory, which emphasizes the importance of shared goals and plans in teamwork. STEAM's key features include the use of joint intentions as the basic building block of teamwork, the ability to monitor team performance and reorganize as needed, and decision-theoretic communication selectivity to reduce communication overheads. The model is implemented in a way that allows agents to reason about their coordination and communication responsibilities, and to anticipate and recover from teamwork failures. STEAM has been tested in three different complex domains: Attack, Transport, and RoboCup synthetic soccer. In the Attack domain, STEAM was used to manage a team of synthetic attack helicopters, where agents had to coordinate and communicate to achieve a common goal. In the Transport domain, STEAM was used to manage a team of synthetic transport helicopters, where agents had to coordinate and communicate to ensure the safe transport of troops. In the RoboCup synthetic soccer domain, STEAM was used to manage a team of synthetic soccer agents, where agents had to coordinate and communicate to achieve a common goal. The results of these experiments showed that STEAM was effective in enabling flexible and coherent teamwork in these complex, dynamic environments. The model was able to handle uncertainties, such as differing views of the environment, unexpected failures, and unexpected opportunities. STEAM's ability to monitor team performance and reorganize as needed was particularly important in these environments. STEAM's implementation is based on the Soar architecture, with a set of about 300 domain-independent Soar rules. Three different teams have been developed based on this operationalization of STEAM. These teams have a complex structure of team-subteam hierarchies and operate in complex environments. The article presents detailed experimental results from these teams, illustrating the benefits of STEAM in their development. STEAM is among just a very few implemented general models of teamwork. Other models include Jennings' joint responsibility framework in the GRATE system and Rich and Sidner's COLLAGEN, which both operate in complex domains. While Section 7 will discuss these in greater detail, STEAM significantly differs from both these frameworks, via its focus on a different (and arguably wider) set of teamwork capabilities that arise in domains with teams of more than two-three agents, with more complex team organizational hierarchies, and with practical emphasis on communication costs.