Adaptive Workflow = Web Services + Agents Paul A. Buhler College of Charleston Department of Computer Science 66 George Street Charleston, SC 29424, USA pbuhler@cs.cofc.edu José M. Vidal University of South Carolina Computer Science and Engineering Columbia, SC 29208, USA vidal@sc.edu Harko Verhagen Department of Computer and System Sciences Stockholm University / KTH Forum 100 16400, Kista, Sweden verhagen@dsv.su.se Abstract Workflow management systems exactly enact business processes described in a process description language. Unfortunately, such strict adherence to the prescribed workflow makes it impossible for the system to adapt to unforeseen circumstances. In this paper we propose that workflow description languages and their associated design tools can be used to specify a multiagent system. Specifically, we advance the idea that the Business Process Execution Language for Web Services can be used as a specification language for expressing the initial social order of a multiagent system, which can then intelligently adapt to changing environmental conditions. keywords: workflow enactment, coordination technology, multiagent systems, process description languages 1. Introduction Advances in Information Technology (IT) are creating opportunities for business enterprises to redesign their information and process management systems. The refinement of service-oriented architectures and the emergence of web-enabled, semantically described services allow us to envision a future where these Web services become the next generation of enterprise components. This new enterprise software vision places new demands on software architectures because they will need to support computing with “dynamically-formed, task-specific, coalitions of distributed autonomous resources” [1, pg 99]. These changes are a logical consequence of the seminal work in coordination technology done by Gelertner. It is now generally accepted that Gelertner was correct when he theorized that computation was orthogonal to coordination [2]. This orthogonality was implied by DeRemer, who wrote in 1976, “Structuring a large collection of modules to form a ‘system’ is an essentially distinct and different intellectual activity from the construction of the individual modules [themselves]” [3]. From these perspectives, a software system is viewed as an ensemble of coordinables and their orchestrated interactions. Coordinables are entities that function as independent units of computation. The coordinated interaction of the computational units produces the desired behavior of the system. Obvious parallels to workflow systems exist; the workflow activities are the coordinables and business processes coordinate their interaction. Leymann asserts that workflow construction can be viewed as a two-level programming problem [4, pg 217]. His view is that the implementation of workflow activities is akin to traditional programming, or programming in the small. Activities encapsulate well-defined functionality that typically involves low-level data access routines and algorithmic processing. In contrast, the building of the workflow’s process model is akin to programming in the large. The process model prescribes coordination rules by providing a means to express the sequencing of the activities and the flow of data amongst them. We advocate the synthesis of Gelertner’s and Leymann’s points of view. We believe that the statements applications = computation + coordination and workflow = activities + processes are equivalent. This paper presents our vision that multi-agent systems are a required ingredient for the flexible enactment of enterprise workflows. Our view can be summarized by the aphorism Adaptive Workflow = Web services + Agents. In this context, the Web services provide the computational resources and the Agents provide the coordination framework. We propose the use of the Business Process Execution Language for Web Services (BPEL4WS) as a specification language for expressing the initial social order of a multiagent system. 2. Adaptive Workflow Enactment Traditionally, workflow management systems have not been designed for dynamic environments requiring