Incorporating intelligent agents into human teams presents many challenges. How should the agents be structured? What roles should they play in the overall team context? Can these roles be adapted during task performance? Is such adaptation beneficial? How can collections of agents be robust and maintain efficient performance in the face of appearance and disappearance (or relocation) of other team members, information sources and communication links? What are effective ways for intelligent assistants to interact with the human team members and with each other so as to increase team effectiveness? What are appropriate measures of agent effectiveness within a team context and team effectiveness of teams consisting of human and machine members? These are some of the many challenges that our research is addressing.

Characteristics distinguishing successful teams from unsuccessful ones include team self-awareness, within-team interdependence, performance monitoring, feedback, clearly communicating intentions, and helping team mates when needed. To contribute to team success, intelligent agents must support these forms of group interaction as well as more task oriented functions. To date technology has focused on a single agent acting as a single user's information gatherer. The single agent approach clearly will not work in a complex team environment where the requirements of large amounts of information, heterogeneity of supported functions, and high performance make the signle agent solution impracticable. To aid in the fast paced, multi user environment of joint mission planning, agents must develop greater capacities for modeling users and situations, preparing and communicating task information, adapting to changes in situation and capabilities of other team members, and providing cues about their own performance and capabilities.

Human factors and cognitive science research have developed detailed methods for representing tasks at the individual level. Team tasks, by contrast, have led to impoverished models such as queueing simulations or models focused on group dynamics rather than task content. However, attention to collaborative and team tasks is growing. Limited research to date has suggested that besides an individual member's task, there is also a team level task that guides the activities of the team members. The interplay among an individual task model, characteristics and individual differences of a team member, the team task model, and the situation guide the team activity and coordination. In current practice, the team task model is represented in documents describing the overall team mission, but primarily it is implicit in the heads of the team members. In teams consisting of humans and intelligent agents, the team model and individual task models must be explicitly and formally represented to enable the machine agents to: (1) be aware of task interdependences that help them identify how they could and should work together, (2) track team member activities (human and machine members) in the context of team tasks, (3) become aware of deviations.

The identification of team and individual tasks, allocation of roles and functions for performing those tasks, and defining task and role models for humans and their intelligent assistants are key activities for our project.

We believe that co-training and mutual adaptation are the key to successfully integrating intelligent assistants into high performance teams. The interactive team debriefings serve both to aid assistants in adapting to their human teammates and to familiarize human team members with the agents' capabilities, rationales and behaviors. As assistants adapt in anticipating particular team members' informational needs they will come to approximate more and more closely the task sharing and cohesion exhibited by members of well trained human teams. Conversely, as intelligent assistants' capabilities, quirks, and defaults become better known to the team the assistants will become more predictable and hence valuable as ready to hand tools which can be wielded with growing flexibility. While simplicity, transparency, and effective explanation facilities seem essential to fostering trust, understanding, and adaptation we hope to develop a more sophisticated model of what makes agents comprehensible to human users to guide the development of new generations of increasingly usable agents.

The human factors literature on teamwork has identified the following dimensions for effective Teamwork:

• Shared Situation Assessment
• Supporting Behaviors
• Team Leadership/Initiative
• Communication