The work presented here is an extension of previous PAR work , which uses natural language input to dynamically alter the behaviors of agents during real-time simulations. In that architecture we were able to give both immediate instructions and conditional instructions to the agents.  However, we were not able to give negative directives or constraints, such as ``Do not stand in front of the car door'' or ``Do not walk on the grass.''  In order to successfully carry out these type of instructions, we need a planner which can dynamically alter an agent's behavior based on constraints imposed by new instructions. 

We extend our current architecture to include a planner and the ablility to keep a record or history of all actions successfully completed by the agent.  This will  also allow us to give conditional, iterative instructions to the agent, such as ``Check every abandoned vehicle.'' The history feature ensures that the agent will correctly check every abandoned vehicle only once.  The planner also allows us to give abstract instructions such as, ``Take cover behind the drum.'' This instruction does not include information about how to take cover behind the drum.  The agent could walk, run, crawl, swim, or do any number of other translatory actions to position itself behind the drum.  Also, the instruction has no information about the path the agent should take.  The planner will provide the information that is needed for animation but not provided by the natural language input.