Building Applications on a Workflow Platform


by David Green

Summary: A workflow can be useful for resolving business issues. Here we'll get acquainted with and investigate the idea of building applications on a workflow platform. A workflow platform supports key workflow concepts and provides a basis for building applications structured using these concepts, including workflow products as hitherto understood. We'll survey a range of applications to explore the necessary characteristics of a workflow platform, which leads to a discussion of the potential benefits of building applications on a workflow platform. We'll also discuss the Windows Workflow Foundation as a means of realizing these benefits in practice.


A Workflow Model
Workflow Contracts
Problem-Resolution Collaboration
Scripted Operations
Rule and Policy
Workflow Platform Value
More Semantic Exploitation
Platform Characteristics
A Common Workflow Runtime
Attacking the Problems
About the Author

The idea of workflow has been around for a long time, and it has been consistently attractive as a way to attack business problems. The problems which a workflow approach has been applied often display three characteristics: The key business value delivered is coordination, for instance, in organizing multiple contributions to the preparation of a quote or driving a document review. Each instance of the business process concerned is of long duration, measured often in days, weeks, or months, rather than minutes. The business process has human participants, who usually contribute most of the work product.

However, only a small proportion of the business problems with these characteristics are solved using a workflow approach. Most commonly, the business process is not recorded as machine-readable information at all. Rather, each of the humans participating in the business process interacts with business systems that are not aware of the semantics of the process as a whole, such as a customer information system, and with other human participants through content-neutral communications channels such as e-mail. Each human participant uses a mental model of their part in the overall business process to determine their behavior.

The attractions of creating a machine-readable model of the business process, that is, a workflow, are not difficult to envisage. Three key benefits that a workflow model can bring are insight, monitoring, and optimization. A set of related workflow models can be used to gain insight into the flow of work through an organization. For monitoring, knowing which individuals are contributing work to which business process is very useful when trying to understand costs and workloads. For optimization, having a model of the work being undertaken, and being able to use the model to interpret behavior, together make it possible to reason about how to optimize the business process.

A Workflow Model

Given these compelling benefits, why haven't workflow models been used more widely? The most likely answer is that the cost of using them has been too high. These costs include product costs, that is, the direct cost of purchasing a workflow product; integration costs, where processes modeled as workflows need to be integrated as part of a larger business system; and standardization costs, where it is difficult for a large organization to standardize on a single workflow technology. Variations in workflow products also mean that skills and model portability are issues.

Let's look at the possibility of addressing these blocking issues by building applications on a workflow platform that is low cost, ubiquitous, uniform, and easily integrated in applications. To be clear, the idea is not to replace workflow products. Rather, the hypothesis is that it is useful to factor out support for some core workflow concepts into a platform on which both workflow products and other applications can be built (see Figure 1).

A workflow is a model, which means it is a machine-readable description of business behavior that is not code. The meaning and benefits of this concept in the context of the value of a workflow platform will be discussed later.

A workflow model describes an organization of work units. For instance, suppose that a document review process specifies that Joe writes the document and then Fred reviews it. Here, the work units are first writing and second reviewing the document, and the organization is that one task must follow the other. This concept is not a radical idea. Code that makes successive calls to two subroutines is a valid example of the concept. The interest lies rather in the forms that this organization takes.

To test the workflow platform hypothesis, we will consider a range of real-world applications and explore the characteristics that a workflow platform should have if it is to prove useful.

A document review process takes as an input parameter a set of [reviewer, role] pairs that describe which humans are involved in the workflow in which roles. Possible values for the role are required, optional, final approver, and owner. The review process then proceeds until all reviewers have performed their assigned roles and notifies the owner of the outcome.

Here, the work items are the document reviews organized by the review process. There are three interesting characteristics to call out, namely, multiple points of interaction, human and automated activity, and the need to handle dynamic change.

Workflow Contracts

The workflow has multiple points of interaction, or contracts. First, there is a contract with a reviewer. This contract involves asking a reviewer to review a document, accepting the verdict and any review comments, and also telling a reviewer that his or her input is no longer required (if the review is canceled, or perhaps if enough reviewers have voted yes). The contract might also allow a reviewer to delegate a review. Then there is a second contract with the final approver, which is a specialization of the reviewer contract. Third, there is a contract with the owner of the review that allows the owner to cancel the review and be notified of the outcome of the review. Finally, there is a contract with the initiator of the review process, who instantiates the review and supplies the required parameters.

It is typical of workflows that they connect multiple parties through a variety of contracts (see Figure 2). The document review workflow is essentially a coordinator, initiated through one contract, that is coordinating a variety of participants through one or more additional contracts.

The document review workflow drives human activity. However, it might also drive automated activities, such as storing versions of the document in a repository as the review progresses. From the point of view of the workflow, there is no essential difference. A workflow can be thought of as communicating, in general, with services through contracts. One special case of a service is another workflow. Another special case is a human. In many ways, a human is the original asynchronous service: one never knows when or if it is going to respond.

A characteristic of this type of workflow is that the participants will ask for changes to the workflow as it executes. For example, a reviewer might delegate a review task to a colleague or share the work involved in a review task with a subordinate.

There are two ways of addressing this requirement. One is to build an understanding of all the possible changes into the workflow. Then, a delegation request becomes just another function of the contract between the workflow and the reviewer. The other possibility is to see change as something separate from the workflow, where change is implemented as an external function that changes the workflow model. In this approach, the result of delegation is a new workflow model identical to one in which the review task was assigned to the delegate from the beginning.

Requesting an additional approval step would add a new approval task to the workflow model, which might well have contained no approval steps at all in its original form. The workflow no longer has to anticipate all possible modifications; at the most it will be concerned with restricting the areas of the model that are subject to change.

Both approaches are useful. Building understanding into a workflow is simple to model and understand. Generalizing operations is more complex to model, but more powerful and agile.

In an extreme but interesting case of the latter approach, the workflow begins execution with little or no content, and the required behavior is added dynamically by the participants in the workflow. Here, the available operations for modifying the workflow become a vocabulary that a user can use to construct the desired behavior as the workflow progresses.

Problem-Resolution Collaboration

To look at a specific example of a problem-resolution collaboration application, consider an inventory shortfall. An assembly line is making a gadget, and the computer indicated that there were enough widgets in stock for the purpose. However, when the stockroom manager went to fetch the widgets for delivery to the assembly line, a shortfall of 10 widgets was discovered.

Collaboration among the stockroom manager, the customer's account manager, the supplies department, and the production manager is required to resolve the problem. Each role in the collaboration may take characteristic actions. The supplies department could order more widgets, perhaps using a different supplier or paying an existing supplier more money for faster delivery. The account manager could go to the customer and request deferred delivery or split the delivery into two parts and bear the extra shipping cost. The production manager could divert assembled gadgets from an order for another customer. The stockroom manager could search his or her physical stock in an attempt to find the missing widgets. Any given action might be performed multiple times.

One obvious constraint is that the collaboration is not completed until the shortfall is resolved by some combination of the previous actions. There will often also be business constraints. For instance, there might be a rule that says deferral of delivery to gold customers is never permitted. Also, the actions will affect each other. For instance, there might be a policy that states that total added cost from corrective action may not exceed 5 percent of original factory cost. Thus, placing an order for accelerated supplies at a higher price might prevent a shipment from being split.

Click here for larger image

Figure 1. A monolithic workflow to stack (Click on the picture for a larger image)

In this case the work items are the actions that the various participants can take as they seek to resolve the inventory shortfall. The organization, however, is not the same as that required in document review. The participants are not dictated to; instead, they choose which actions to perform and when to perform them. However, these choices are constrained by the organization of the workflow, which has two aspects: 1) The actions are focused on achieving a goal; in this case, resolving the inventory shortfall. A bounded collaboration space is created when the problem resolution starts, and is not closed until the goal has been reached. 2) The participants are not free to perform arbitrary actions. Instead, the available actions are determined by the role the participant is performing and the state of the collaboration. The set of actions available is determined by policies related to the goal and global policies such as the restriction on shorting gold customers. The actions available vary as the collaboration progresses.

Click here for larger image

Figure 2. A contract diagram for the document review application (Click on the picture for a larger image)

The experience of the participant is no longer that of performing assigned tasks. Instead, a participant queries for the actions currently available to him or her, performs none or one of these actions, and then repeats the cycle.

The main new requirement here, therefore, is for a form of organization of work items that is essentially data state and goal driven. There is also a requirement to support a query/act-style of contract with a workflow participant.

Scripted Operations

Scripted operations are simply a set of operations that are composed using a script. An example might be a desktop tool that allows a user to define and execute a series of common tasks, such as copying files and annotating them.

It would be unusual to consider using a typical workflow product for this purpose. However, it does fit the workflow platform pattern of a set of work units organized by a model. In this case the model is a sequence, perhaps with support for looping and conditional execution. Therefore, if a workflow platform were sufficiently low cost and ubiquitous, it would be possible to consider applying it to this sort of problem. Would doing so add any value?

One feature of scripted operations that is not addressed by their typical implementations today is the question of data flow. It is common for the data required by one operation to be the output of some previous operation, but this information is not typically modeled in the script. Thus, a user assembling tasks using a desktop tool might not be told when creating a script that the prerequisite data for a task hasn't been supplied, and would only discover the error when running the script. A workflow model that can describe these data dependencies would add clear value for script authors.

One approach is simply to include data flow constructs in the workflow model. It is highly arguable that the basic workflow model needs to include basic structural features such as sequences, conditions, and loops; but it is not clear that data flow is sufficiently universal to be represented by first-class elements of the model.

An alternative approach is to layer support for data flow on top of an extensible, basic workflow. A workflow model that can be enriched with abstractions appropriate to a variety of problem domains fits well with the notion of a workflow platform. This approach avoids both the complexity created by including in the base model a large variety of semantic constructs specialized for different problems and also the limitations imposed by limiting the workflow model to a fixed set of constructs.

Now let's look at a guided user application. One example is an interactive voice response (IVR) system, and another is a call center system guiding telephone operators through support or sales scripts. The essence of these applications is to guide users through the series of operations needed to achieve their goal. The organization of these operations is typically used to drive the presentation to the user, whether this is generated speech or a set of enabled and disabled command buttons on a form.

A characteristic of this type of application is that the workflow is the most frequently changed part of the application. Also, the business sponsors of the system are often heavily involved in specifying the changes, making it important to provide a way for IT staff and business personnel to communicate clearly and efficiently about the changes. A workflow model that expresses the core business purpose of the application, stripped of any irrelevant technical material, is an effective way to achieve this communication.

These applications also require flexibility within the workflow structure. In an IVR application the user will typically be heavily constrained, moving through a hierarchically structured set of menus. However, there will also be escape commands—for example, "return to root menu" or "skip out of current subtree."

A call center application will have more flexibility than an IVR application, changing the options offered to the user in response to the state of an order or in response to the input from a customer, such as skipping sales prompts if the customer starts to react negatively.

This sort of application requires support for a mix of organizations of work items, combining sequences, loops, and conditions with jumps from one state to another, and also the kind of data-driven behavior seen in problem-resolution collaboration.

Rule and Policy

As discussed previously, one way in which the workflow approach can deliver value is by isolating the focus of change in an application. Often, this focus is on the way in which the work items are structured, but in some applications the focus of change is on expressions tied to a relatively slow-changing structure.

An example of this focus is an insurance policy quotation system, where a set of frequently changing calculations is used to drive decision making in the quotation process. The requirement is for the workflow to model these expressions, which has two key benefits: First, the testing and deployment costs are much lower than those that would typically be incurred if the expressions were written as code, since the model provides a strong sandbox restricting the scope of possible changes. Second, the changes can be made by personnel who understand the business significance of the expressions but do not have the skills to understand the technical code in which expressions written as code would inevitably need to be embedded.

The Model-View-Controller (MVC) pattern often is used to wire a UI to an underlying object model (see Figure 3). The model represents the behavior of the system, independent of any particular UI representation. The controller is a part of the UI layer that is used to map the events generated by the UI into the method invocations required to drive the model. The UI itself is thus not polluted by any assumptions about the underlying model.

The workflows considered so far, viewed from this standpoint, all fall into the category of Models in the MVC sense. However, the controller can also be seen as a workflow. The work items it organizes are the methods provided by Model objects. The controller also interacts with the UI and the model through well-defined contracts. A model of this kind is often termed a page flow.

As with scripted operations, page flow would not today be implemented using a typical workflow product. There are two reasons to consider building a page flow using a workflow platform. First, a model readily can be represented visually, helping developers and analysts to express and communicate the required behavior. Second, if the page flow is frequently changing, then the abstraction of the page flow as a model improves agility.

There are two main requirements if this problem is to be addressed using a workflow platform. The workflow runtime must be lightweight, since a page flow may be running within a small application on a desktop, and the contracts supported must include the event-based contract characteristic of UIs, as well as the object-method contracts exposed by the Model.

Now let's look at a test record/replay application example. The intent of this final example is to test the limits of the applicability of the workflow platform hypothesis.

Click here for larger image

Figure 3. An MVC application (Click on the picture for a larger image)

The application here is a tool for testing applications built as a set of services. The tool uses an interception mechanism to record all the interaction between services that occur during manual performance of a test case for the application. This recording can then be replayed. During replay, externally sourced messages are generated without manual intervention, and messages between the set of services that make up the application are checked for sequence and content against the original recording.

The workflow is the test case, organizing the work units that are the participating services. The workflow is both active, in that it simulates the behavior of externally sourced messages, and passive, in that it monitors the interactions between services.

A unique feature of this application is that the workflow is written, not by a developer or a user, but by a program, as part of the act of recording a test case. Workflow model creation must be fully programmable. There are also requirements for extensibility and dynamic update.

Extensibility is required because the structural semantics are rich. For instance, just because two messages arrived at a service one after the other in the recording, there is no necessary implication that this order needs to be preserved in a replay. If there is no causal dependency between the messages, then a replay that reverses the order of the messages is correct. So the semantics of sequence in the model used to record the test cases need to include a notion of causality, which is not likely to be a feature of the core workflow model of sequence.

Dynamic update is required because human interaction with the model will occur during replay. Discrepancies discovered during replay between recorded and observed behavior are popped up to a tester. If the discrepancy is because a message includes a timestamp that varies from run to run, then the tester would update the model to mark the field "don't care." If the discrepancy occurs in a regression test because the software has changed, then the tester might approve the change and update the test to expect the new behavior in all subsequent runs.

Workflow Platform Value

A workflow platform does not, by definition, have the full set of features offered by today's typical workflow products. Rather, the workflow platform considered here focuses on supporting the concept of a workflow as a model of the organization of work items. We have seen that this idea of an organization of work items is indeed applicable across a broad range of applications, but the fact that a workflow platform can be used doesn't mean that it should be used.

Click here for larger image

Figure 4. The document review implementation schematic (Click on the picture for a larger image)

Two questions must be asked: What additional value is derived from the workflow platform approach? And, is this approach practical? This value of the workflow platform approach must come from the expression of the organization of work as a model, which we'll discuss later. Let's summarize the characteristics that a practical and effective workflow platform must display.

To demonstrate how a model differs from code, this code snippet is a valid workflow by the definition used here, that is, an organization of work units.

public void HandleLoanRequest (
    string customerID,
    Application app)
    if (CheckCredit(
        customerId, app.Amount))
        MakeOffer (customerId, app);

And, in a sense, it is a model. It is possible to parse this code and build a CodeDOM tree that represents it.

However, the semantics of the resulting model are so general as to be opaque. It is possible to tell that the code contains function invocations, but it isn't too easy to distinguish a function that represents the invocation of a work item from a function that converts integers to strings. A workflow model explicitly distinguishes these ideas. Typically, a specialized model element is used to represent the invocation of a work item, and conversion functions cannot be expressed directly in the model at all. A workflow model, then, is one in which its graph is built from elements that are meaningful in the workflow domain. The semantic richness of such a model can be exploited in several ways.

Visualization. Visual representation of the model—typically in graphical form—is useful for the developer, during both development and maintenance, and also for workflow users who want to know why they have been assigned a given task or the IT operations worker who wants to understand what a misbehaving application should be doing.

Insight. The workflow model is amenable to programmatic access for a variety of purposes. An example is static analysis to determine dependencies and flow of work across a set of cooperating workflows or using the model to drive a simulation that predicts the workloads that will be generated by a new version of a process.

Expressiveness. The specialization of the workflow model to the workflow domain means that characteristic problems can be expressed more quickly and compactly. It is a domain-specific language (DSL), specialized to support characteristic problems. Consider a document review process where three positive votes out of five reviews mean that the document is good, and any outstanding reviews can be canceled. This process is quite difficult to code, but a workflow model can supply out-of-the-box constructions that address such problems.

More Semantic Exploitation

As we have seen in the scripted operations application discussion, extending the workflow model to further specialize the out-of-the-box model language is a very powerful technique for delivered additional value. An example is the creation of a language intended for end users, as in the document review conducted using an improvised definition of the review process that was discussed previously.

Execution. The specialization of the model makes it possible to add run-time support for common problems. A good example is long-running state. Of the applications discussed here, management of long-running state is required for the document review process, problem-resolution collaboration, and guided user applications. The workflow platform runtime can solve such difficult problems once, using simple expressive model elements to control a common capability and freeing up the developer to focus on the business problem.

Monitoring. The existence of a model makes it possible to produce an event stream with a meaningful semantic without any additional developer effort. Of the applications described here, this event stream is useful in the document review, problem-resolution collaboration, test record/replay, and guided user applications. The event stream can be used to monitor instances of workflows or build aggregate views of the state of a large number of workflow instances. The standardization of the event stream makes it much easier to build such aggregate views across workflows that were developed independently of each other.

Another powerful idea is the presentation of errors using a business semantic. Often, a technical failure such as the nondelivery of a message leads to escalation to a technical expert because the significance of the failure is unclear without specialist investigation. If the error can be mapped to a workflow model—so that it is clear that the error concerns a noncritical change notification, for instance—then escalation can be restricted to cases where it is necessary.

Composition. If an application is factored into work units, then these work units, with their well-understood interfaces, can be reused by other workflows. Workflows themselves also define work units that can also be used by other workflows.

Customization. Suppose that an ISV ships a workflow, which is customized by a VAR, and then again by a customer. Reapplying these customizations when the ISV ships a new base version is a challenging maintenance problem. The use of a shared, well-understood model for the workflow makes the consequent three-way merges much more tractable. Customization and composition together enable ecosystems where definitions of work and flow become shared or traded artifacts.

Manipulation. As we have seen in the discussions of the document review and test record/replay applications, there are often requirements to invent or modify workflows on the fly. This modification cannot be done securely if changing code is required. Using a model makes possible dynamic manipulation that is both controllable and comprehensible.

These benefits make a compelling list, and it demonstrates clearly that the description of an organization of work items as a model has a lot to offer.

Platform Characteristics

There must be support for basic structural concepts like sequences, conditions, and loops. However, there also needs to be support for data-driven approaches to deal with the less-structured organizations that appear in applications like problem-resolution collaboration and guided user.

It is also important to allow new semantic elements to be added to create rich, specialized languages such as the data flow-aware composition in scripted operations. Adding new semantic elements might go so far as to require the redefinition of such fundamental ideas as sequence—for example, in the test record/replay application.

The workflow must also be able to communicate in a rich variety of ways. Workflows respond to UI events, drive different types of services (human, programmatic, other workflows), and support queries over the current state of their contracts—for instance, when determining the actions available to an actor in a problem-resolution collaboration application.

If the workflow platform is to be used in all the applications where it adds value, such as MVC, then it must be lightweight. Equally, it needs to address the scale and performance requirements implied by applications such as document review.

In addition, the workflow model itself must be fully programmable, which includes model creation—such as in the test record/replay application—and dynamic model update to support unanticipated change, as in both the document review and test record/replay applications.

Now let's look at the realization of these required characteristics in the Windows Workflow Foundation (WF). Thus far we have effectively recapitulated the thinking that drove the development of WF. WF as a realization of these concepts is the means by which this value can be translated into delivered solutions.

WF implements the idea of workflow as an organization of work items, abstracted away from the related ideas with which it has been coupled in traditional workflow products. The abstractions fall under three main categories: design and visualization, hosting, and semantics.

Design and visualization. A workflow in WF is a tree of work items (called activities). This tree can be manipulated directly as an object model. A designer is provided, but its use is not mandated. It is possible to create new designers specialized to particular user communities or to particular organizations of work items. It is also possible to specialize the provided designer, which can be used not only within Visual Studio but from within an arbitrary hosting application.

Hosting. The WF runtime is sufficiently lightweight to be hosted in a client context such as a controller in a rich-client application shell. It is also performant enough to scale when embedded in a server host, such as the SharePoint Server delivered by Office 2007. The WF runtime's expectations of its host are abstracted as provider interfaces for services such as threading, transactions, persistence, and communications. Useful provider implementations are supplied out of the box, but they may be substituted as required.

Semantics. Different problems respond to different model semantics. WF supports three main styles of workflow out of the box: flow, state machine, and data driven. Flow is optimal for applications where the workflow is in control such as the scripted operations example. State machine is best when the workflow is driven by external events, as in the MVC or guided user applications. A data-driven approach is suited to applications where actions depend on state, as in problem-resolution collaboration.

These semantics can be extended by building custom activities to create a domain-specific vocabulary for use in any of these styles. However, since the structure of a workflow is itself expressed as a set of activities, the same approach can be used to define new styles, and entirely novel semantics, if required.

A Common Workflow Runtime

The focus of the WF runtime is to deliver facilities required by any workflow, and therefore avoid the need to re-implement them time and again in different applications, but without compromising the flexibility of the workflow abstraction. These common facilities fall into four main categories: activity scheduling, transactions and long-running state, exceptions and compensation, and communications. Let's look at each in more detail.

Activity scheduling. The WF runtime defines an activity protocol that all work items implement. This protocol defines the basic activity life cycle (initialized, executing, and closed) and the additional states needed to handle exceptions (faulted, canceling, and compensating). This definition enables the WF runtime to provide work scheduling for all workflows.

Transactions and long-running state. The WF runtime supports the execution of ACID transactions. These transactions are particularly useful for maintaining consistency across workflow state and external state such as application and message state. However, ACID transactions are not suitable for managing long-running state because of their resource and locking implications. The WF runtime implements a broader checkpoint-and-recovery mechanism to handle long-running state. From this point of view, ACID transactions become units of execution within a larger framework. The developer needs not do any work to get the benefit of WF's support for long-running state, as it is default behavior. However, if more detailed control is required, a set of simple model elements are supplied for the purpose.

Exceptions and compensation. The familiar idea of throw-try-catch exceptions is supported by the WF runtime and represented in the out-of-the-box workflow model. However, the WF runtime also supports a broader view of fault handling that includes the idea of compensation for successfully completed transactional units.

Communications. As we have seen, workflows need to communicate in a variety of ways, which is reflected in the WF, that supports communication through .NET method, event interfaces, and Web service interfaces. Support for Windows Communication Framework will also be made available in the future. Thus, WF does indeed realize the workflow-platform approach proposed here.

Figure 4 illustrates the high-level implementation schematic of the document review application and how all of the foregoing comes together. An implementation uses SharePoint as the workflow host. The WF runtime uses the default scheduling service provided out of the box with WF. However, the default persistence and communications services are replaced with implementations specialized for the SharePoint host. The persistence service stores long-running workflow state in the SharePoint database, and the communications service makes the rich-user interaction facilities of SharePoint available to the workflow. Both of these services are in fact delivered out of the box with Microsoft Office 2007.

Three sorts of activities are used to define the document review workflow itself. First, out-of-the-box WF activities are used to provide structural elements such as If-Else and While. Second, activities provided as part of Office 2007 are used to access the user communication services of SharePoint. Third, custom activities are used to implement organization-specific semantics for forwarding and delegation in a standard and reusable way. The WF designer is used as a means to define the workflow and also provide diagrammatic representations of the state of a document review workflow instance to the workflow owner.

Attacking the Problems

In summary, the workflow platform supports an abstraction of the ideas that have made workflow products an attractive attack on business problems. It does not replace today's workflow products, however. Rather, it factors them into platform and superstructure.

The workflow platform embodies two key ideas: a workflow is an organization of work units, and a workflow is a model, that is, a machine-readable description other than code. These ideas are valuable in a broad range of applications, both within and beyond the problem domain addressed by typical workflow products. Such a workflow platform is most useful if it is low cost and ubiquitous.

The principal benefits delivered arise from the expression of an organization of work items as a model, which has several advantages over a representation in code:

  • Transparency. The business purposes of the system are clear, allowing users and IT staff to communicate effectively about the desired behavior and IT staff coming onto the project to get up to speed quickly.
  • Isolation of change. The areas of the application most likely to change are expressed as workflow rather than code. By isolating the rapidly moving parts of the application, changes can be made more reliably.
  • Agility. The bottom line of all these benefits is business agility. If business users can understand the system, developers can get up to speed quickly, and the risks associated with change are minimized. Then the system may be termed agile.

A broadly useful workflow platform must have these characteristics: define a core workflow model as a standard that is extensible and fully programmable at design time and runtime, be able to communicate in a rich variety of ways, be lightweight and embeddable, and be able to scale and perform well in high-volume environments. WF is a product that displays all of these characteristics. As a component of WinFx and a part of the Windows platform, WF is also low cost and ubiquitous.

The concept of a workflow platform, as described here, is an appropriate and customizable model with multiple benefits for a wide range of applications, and it is fully realized in WF.

About the Author

David Green joined IBM as a developer in the Hursley labs in 1977. Since then, he's moved up and down the software supply chain, working for a newspaper company, American Express, Siemens Nixdorf, and a major UK bank in a variety of technical, presales, and post-sales roles. The theme that runs through all of his experience is building applications for the day-to-day business world, thinking it was far more difficult than it should be to create great business solutions, and trying to create approaches and tools to do something about it. David has spent the last two years at Microsoft working on Windows Workflow Foundation, which he believes is a significant addition to the application builder's armory.

This article was published in the Architecture Journal, a print and online publication produced by Microsoft. For more articles from this publication, please visit the Architecture Journal website.

© Microsoft Corporation. All rights reserved.