Extensive research in autonomous agents has studied actionselection, i,e. the problem of having an agent at any point in time choose actions that will best fulfill its goals. However, for many applications where a human interacts with the ngent, it is not enough for the agent to 'do the right thing;' it must also do it in the right way, i.e. so the user can understand what the agent is doing. Tom Porter terms this problem the 'action-expression' problem: what should the agent do at any point in

more »

... order to best communicate its goals and activities to the user? Current agent architectures often have difficulty with action-expression because design of the agent is focused on internal problem-solving rather than external effect. Behaviorbased agents in particular tend to jump from behavior to behavior according to whatever best fulfills their internal needs, which can confuse a user trying to find a common thread in the agent's activities. The system described here, the ,!hpreasivator, is based on a philosophy that what matters is not the internally-defined code as understood by the designer, but the impression the agent makes on the user. The Expressivator's focus is on reducing the apparent randomness of agent behavior choice by adding transition be-IIauior8, special behaviors that function to explain to the user the agent's motivations in changing from one activity to another. In addition, the Expressivator offers a sign-mana~cmcnt suatem that keeps track of the visible signs the agent's behavior has produced, allowing the agent to make decisions based not only on its internal idea of what it is doing, but also on the likely user perception of its behaviors. It bootstraps on the advantages of behavior-based systems, including reactivity, interruptability, and modularity, while allowing the agent-builder to design agents that explicitly communicate their goals and intentions to the user.