Developing a computational framework for capturing the
shared mental model among members of effective teams is
a challenging and critical issue for applications ranging
from team training to supporting teamwork. The CAST
architecture supports flexibility in its teamwork knowledge
specification and in dynamic role selection at run time. At
the same time, it leverages shared knowledge about the
structure and the process of a team to reason about
information needs of teammates efficiently. We believe the
CAST architecture achieves a reasonable tradeoff between
the flexibility and the efficiency for simulating proactive
information exchange among teammates. While our
experimental results demonstrate anticipated benefits of
CAST, it also reveals some limitations of the current CAST
implementation. For example, we plan to extend the role
selection method for finding backups when an agent dies or
becomes non-functional. With such an extension, we hope
to simulate more complex teamwork behavior.