161 Zwitserloot

Software agent technology generally is defined as the area that deals with writing software in such a way that it is autonomous. In this definition, the word autonomous indicates that the software has the ability to react to changes in its environment in a way that it can continue to perform its intended job. Specifically, changes in its input channels, its output channels, and the changes in or the addition or removal of other agent software should cause the agent to change its own behavior in order to function properly in the new environment. In other words, the term software agent refers to the fact that a certain piece of software likely will be able to run more reliably without user intervention in a changing environment compared to similar software designed without the software agent paradigm in mind. This definition is quite broad; for example, an alarm clock that automatically accounts for daylight savings time could be said to be autonomous in this property; a change in its environment (namely, the arrival of daylight savings time) causes the software running the clock to adjust the time it displays to the user by one hour, preserving, in the process, its intended function—displaying the current time. A more detailed description of agent technology is available from Russel and Norvig (2003). The autonomous nature of software agents makes them the perfect candidate for operating in an environment where the available software continually changes. Generally, this type of technology is referred to as multi-agent systems (MAS). In the case of MAS, the various agents running on the system adapt and account for the other agents available in the system that are relevant to its own operation in some way. For example, MAS-aware agents often are envisioned to have a way of negotiating for the use of a scarce resource with other agents. An obvious start for developing MAS is to decide on a common set of rules to which each agent will adhere, and on an appropriate communication standard. These requirements force the need for an underlying piece of software called an agent framework. This framework hosts the agents, is responsible for ensuring that the agents keep to the rules that apply to the situation, and streamlines communication between the agents themselves and external sensors and actuators (in essence, input and output, respectively). This paper will go into more detail regarding the advantages of MAS and agent frameworks, the nature and properties of agent frameworks, a selection of frameworks available at the moment, and attempts to draw some conclusions and best practices by analyzing the currently available framework technology.