You are on page 1of 11

November 2004

An Introduction to Agile Software Development
by Victor Szalvay, co-founder
Danube Technologies, Inc.
Web site:
Web log:
12011 Bel-Red Rd. Suite 201
Bellevue, WA 98005
(425) 688-0888, ext. 812

This paper is an introduction to the Agile school of software development, and is primarily
targeted at IT managers and CXOs with an interest in improving development productivity.
What is Agile? How can Agile help improve my organization? First, I introduce the two broad
schools of thought when it comes to software development: traditional sequential, a.k.a. “the
waterfall method”, and iterative methods of which Agile is a subset. My objective is to
demonstrate the short-comings of the waterfall approach while providing a solution in iterative,
and more specifically, Agile methods.

Part I – Shortcomings of Traditional Waterfall Approach
The essence of waterfall software development is that complex software systems can be built in a
sequential, phase-wise manner where all of the requirements are gathered at the beginning, all of
the design is completed next, and finally the master design is implemented into production
quality software. This approach holds that complex systems can be built in a single pass, without
going back and revisiting requirements or design ideas in light of changing business or
technology conditions. It was first introduced in an article written by Winston Royce in 1970,
primarily intended for use in government projects1.
Waterfall equates software development to a production line conveyor belt.
“Requirements analysts” compile the system specifications until they pass the finished
requirements specification document to “software designers” who plan the software system and
create diagrams documenting how the code should be written. The design diagrams are then
passed to the “developers” who implement the code from the design (See Figure 1).
Under the waterfall approach, traditional IT managers have made valiant efforts to craft
and adhere to large-scale development plans. These plans are typically laid out in advance of
development projects using Gantt or PERT charts to map detailed tasks and dependencies for
each member of the development group months or years down the line. However, studies of past
software projects show that only 9% to 16% are considered on-time and on-budget2. In this
article, I attempt to summarize current thinking among computer scientists on why waterfall fails
in so many cases. I also explore a leading alternative to waterfall: “Agile” methods that focus on

Copyright © 2004 Danube Technologies, Inc. All rights reserved.

Analysts slave for weeks or months compiling everything they can gleam about the proposed system into comprehensive “Software Requirements Specification” (SRS) documents. One of the biggest problems with waterfall is that it assumes that all project requirements can be accurately gathered at the beginning of the project. The cost of change in a waterfall project increases exponentially over time because the developer is forced to make any and all project decisions at the beginning of the project. it is inevitable that attempts at up-front requirements specification will leave out some very important details simply because the stakeholders cannot tell developers everything about the system at the beginning of the project. the intricately planned design can be affected dramatically. consider the areas that must be addressed: business rules and exceptions. What if your business needs are still emerging and certain aspects of the system are rapidly changing or cannot be defined yet? Business climates and objectives often change Copyright © 2004 Danube Technologies. user interface standards. Requirements define what developers are to build.3 This means that the requirements typically change outside of the requirements phase in the form of “change orders”. Do you think you could provide every last detail the developers need to know right off the bat? I have yet to encounter such a customer and I am hard pressed to think I ever will. Once finished.999% of the time. Figure 1 Traditional Methods: sequential phased approach Project 1. In fact. Deploy Analysis Design Phase Up-front Requirements Analysis What are requirements? From the stakeholder’s perspective. design. In Figure 1. which will in turn affect any implementation and test strategies. Test 5. browser or OS support. the system must have a web site with e-commerce capability that can handle 10. the first block represents the requirements analysis phase of a software development project. As a start. user roles and restrictions. Architecture & 3. and testing continue throughout the project lifecycle. By virtue of a requirements change. the requirements are the features and specifications of the system. scalability and concurrent user support. For example.000 purchases per hour. implementation. and in many waterfall projects this can be very costly. Inc. . or the system must be accessible 99. Requirements 2. All rights reserved. Imagine a scenario where you engage a software group to build a critical software system. the SRS is sent over the fence to the designers while the requirements analysts go to work on the next project. Code 4. Page 2 incremental and iterative development where requirements.

functional software. the requirements are solicited from the user and some time later the finished product is presented to the user.g. Can you afford to lock your business into a rigid long-term project where the cost of change grows exponentially? For example. It is therefore fair to say that software development is more akin to new product research and development than it is to assembly-line style manufacturing. But is it possible Copyright © 2004 Danube Technologies. Waterfall is an “over the fence” approach. All rights reserved. If I tried to close my eyes and draw the same picture. and contingency. especially in today’s age of instant information. . organizations developing software must realize variables exist that are largely outside of management control. Markets are forcing the software development community to respond with flexible development plans that flatten the cost of change. the types of questions that would appear on the test were unknown when we started development. Software development is innovation. waterfall projects allocate copious effort detailing every possible risk. Software Development is more like New Product Development than Manufacturing Software development is a highly complex field with countless variables impacting the system. All software systems are imperfect because they cannot be built with mathematical or physical certainty.). But this is what waterfall asks customers to do: specify the entire system without having a chance to periodically see the progress and make adjustments to the requirements as needed. cookie- cutter solutions4. changing. each foray into a development project presents new and difficult challenges that cannot be overcome with one-size-fits-all. I often use a “drawing” analogy to help explain this effect. however. it would prove far less successful. database platforms. a national test preparation organization commissioned my company to build a simulator for an upcoming standardized test. The “I’ll Know it When I See It” (IKIWISI) law says that software development customers can better describe what they really want after seeing and trying working. has no laws or clear certainties on which to build. and those systems that act as building blocks contain bugs and cannot be relied on with certainty.. Inc. This is entirely unnatural because customers find it difficult to specify software perfectly without seeing it evolve and progress. when I draw a picture I need to see the drawing as I progress. In fact. Since the test itself had not been released yet. People need to see and feel something before they really know what they want. Bridge building relies on physical and mathematical laws. etc. The waterfall methodology assumes that up-front planning is enough to take into account all variables that could impact the development process. mitigation plan. programming languages. Also consider that the building blocks of software projects is usually other software systems (e. Although I’m a terrible artist. Page 3 rapidly. The problem of undefined. As a result. and artistry. and emerging requirements presents a very large challenge to waterfall projects because by definition all requirements must be captured up-front at the risk of costly changes later. discovery. But the system had to be done shortly after the tests were released. Because the foundations of software development are inherently unstable and unreliable. Software development. software is almost always flawed or sub-optimized.

HP. Can research really be relegated to a series of steps that when performed in sequence result in a new product? If this formulaic approach were adequate. or iterations. Following the lead of Japanese auto makers. when used sequentially. medical researchers could simply plug variables into equations to discover new medicines. waterfall-style approaches to new product development were effectively abandoned outside the software development industry. Each Copyright © 2004 Danube Technologies. For example. The inherent uncertainty and complexity in all software projects requires an adaptive development plan to cope with uncertainty and a high number of unknown variables. This concept of iterative development hails from the “lean development” era of the 1980s described above where Japanese auto makers made tremendous efficiency and innovation increases simply by removing the phased. since the late 1970s product development companies lead by Toyota. 3M. As new requirements surface and as the scope changes. the benefits of which other new product development industries have been reaping for decades. 6 The results were a dramatic improvement in cost and development time to market and ultimately lead to the popular rise of “lean development” and “just-in-time manufacturing”. Inc. It’s ironic that a cutting edge technology field like software is so far behind more traditional engineering fields in terms of development methods. Almost no software system is so simple that the development can be entirely scripted from beginning to end. Fujitsu. Honda. The development lifecycle is cut up into increments or “iterations” and each iteration touches on each of the traditional “phases” of development.5 Waterfall therefore equates software development to an assembly line. in the 1990s sequential. and NEC. Part II . sequential approach and implementing an iterative approach. and the result is always Z. The first step is X. holistic approach where the traditional phases of development overlap throughout the product lifecycle. . IID processes continually capture the requirements iteration after iteration Interestingly. the second is Y. Page 4 to predict any and all variables that could possibly affect a software project? The empirical answer is “no” considering the limited success of waterfall projects.7 But longstanding insistence from IT managers to categorize software development as a straightforward assembly line progression has kept the software industry from evolving to better methods. Winston Royce (of waterfall process fame) later noted that his ideas were incorrectly interpreted and that a “single pass” framework would never work (his article actually advocates at least a second pass).Iterative and Agile methods Incremental and Iterative Development The simple ability to revisit the “phases” of development dramatically improves project efficiency. over a project lifecycle to properly address complexities and risk factors. On the contrary. result in a successful project each time. The idea of revisiting phases over and over is called “incremental and iterative development” (IID). Canon. with IID requirements is an ongoing process that is periodically revisited. All rights reserved. defined processes can be established that.8 IID allows for multiple “passes”. supplanted the sequential “Phased Program Planning” (PPP) approach to new product development with a flexible. where prototypes were developed for short-term milestones (see Figure 2).

the requirements. like rapid production and minimum up-front design 15. a concise summary of Agile values. and Agile methods. 1986. Agile ideals can be applied to reduce the cost of change throughout the software lifecycle even if the cost of change is not perfectly flat. Inc. All rights reserved. Harvard Business Rev. Agile methods promote an iterative mechanism for producing software. 137-146. Nevertheless. Agile visionary Kent Beck challenged the traditional cost of change curve evidenced by Barry Boehm11 over twenty years ago. design. Nonaka. To accomplish this “flatter” cost of change curve. Page 5 phase was actually a layer that continued throughout the entire development lifecycle. the success of lean development has influenced a broad class of “iterative” software methods including the Unified Process. Takeuchi and I. Spiral. Rather than focusing a lot of effort on big up front design analysis. Agile methods: Embracing Change Agile methods stress productivity and values over heavy-weight process overhead and artifacts. and implementation cycle was revisited for each short-term milestone. "The New New Product Development Game". and they further increase the iterative nature of the software lifecycle by tightening design-code-test loop to at least once a day (if not much more frequently) as opposed to once per iteration. . pp. This “concurrent” development approach created an atmosphere of trial-and-error experimentation and learning that ultimately broke down the status quo and led to efficient innovation.. Agile methods promote a number of engineering practices that enable cost effective change. Figure 2 Iterative approach: Overlapping phases of development Phase 1 2 3 4 5 6 Source: Adapted from H. “Test driven development” is a quality-first approach where developer tests (called unit tests) are written prior to the functional code itself. small increments of functional code are produced according to immediate business need. Beck’s idealistic “flat” cost of change curve has since been revised and softened by Alister Cockburn 13 and Scott Ambler 14 to reflect modern corporate realities. The Agile Manifesto10. Beck’s model espouses that the cost of change can be inexpensive even late in the project lifecycle while maintaining or increasing system quality12. It is the role of the automated test suite built around the rapidly evolving code to act as a harness that Copyright © 2004 Danube Technologies. Evo. was written and signed in 2001 although Agile methods have existed since the early 90s. Author and speaker Martin Fowler describes testing and continuous integration as the “enabling” Agile practices that allow for the advantages gained.9 Although direct analogies between industries are never seamless. Jan.

Figure 3: Cost of Change Curves Cost of Change Boehm Cockburn/Ambler Beck Project Lifecycle Progress Object technology and modern integrated development environments (IDEs) boasting built-in testing and refactoring mechanisms negate the expensive Boehm cost of change curve and allow for the cheap change. Inc. there cannot be a single exhaustive library of defined processes to handle every situation that could possibly surface during a software project. All rights reserved. even late in the project lifecycle. a popular Agile project management method. project plans are continuously inspected and adapted based on the empirical reality of the project. for example. Page 6 allows developers to make aggressive code changes without fear of undetected regression failure. the manufacturing industry has long known that certain chemical processes. . Agile project management approaches balance the four variables in software development while keeping in mind the limits associated with new product development. Risk factors and emerging requirements complicate software development to a point where defined processes fall short. an empirical or adaptive management approach is employed to measure and adjust the chemical process periodically to achieve the desired outcome.16 As a result. In fact. Scrum holds that straightforward defined processes alone cannot be used to effectively manage complex and dynamic software projects. Instead. are too difficult to script and define. Although it has been attempted in the past. In Copyright © 2004 Danube Technologies. changing software projects. introduced the concept of empirical process control for the management of complex. Agile Project Management: Empirical Process Scrum. in the Scrum process.

In reality. middle and upper management often assumes that all four of these factors could be dictated to the development team under the waterfall approach. Although cross-industry analysis can be tenuous. organizations can reduce risks by leaving options open to decide at a better time when more accurate information is available.17 Because software development is often considered a sequential. the concept of frequent or continuous integration keep software development teams synchronized. Agile methods encourage delaying irreversible decisions until the last responsible moment. Customers may not want to make decisions about the system until they have more information. thereby reducing the risks associated with large integrations at the tail end of projects. Consider for example the “small batch” principle: things produced in smaller batches are of higher quality and efficiency because the feedback loop is short. linear process.20 Agile Requirements: A Focus on Business ROI Agile projects avoid “up-front” requirements gathering for the reasons stated above: customers cannot effectively produce all requirements in high enough detail for implementation to occur at the beginning of a project. These factors are interconnected. controls can be adjusted more frequently. § Quality – Cut corners by reducing quality. Page 7 software development there are four broad control factors. and the development process dictates the fourth. Lean Thinking Another effective way to analyze how Agile methods increase efficiencies is to apply lean manufacturing principles to software development. Third. Teams can work independently for a while but the code base never diverges for long periods of time. Rather than locking into decisions at the beginning of a project. . management can pick values for three of the four factors at most. The frequent demonstration and release of software common in Agile approaches Copyright © 2004 Danube Technologies. § Cost – or Effort. § Requirements – The scope of the work that needs to be done can be increased or decreased to affect the project. All rights reserved.18 The highly complex and uncertain nature of software development makes this expectation of full control unrealistic. Second. Available money impacts the amount of effort put into the system. and resources are utilized efficiently to avoid “queuing” (see “queuing theory” and the theory of constraints). However software development cannot be described by a simple linear process because it cannot be predicted accurately in advance. Inc. § Schedule – A software project is impacted as the timeline is changed. Agile methods have their conceptual roots in the Japanese manufacturing productivity boom of the 1980s 19. when one changes at least one other factor must also change. Many software development organizations that implement Agile software development are finding they get something they never expected: options. Agile values a high visibility and customer involvement. It is therefore unreasonable to assume that management can control all four of these factors.

in our experience often the bottleneck in the development process has been the lack of availability of customer domain experts for detailed requirements analysis. For example. a Project Burndown Chart can be utilized to estimate the eventual conclusion of an estimated backlog of work. the development team must bring technical risks to the customer. IID is ideally suited then to take on bite-sized chunks of requirements that the customer can easily digest. An iterative approach allows customers to delay decisions as well. Agile helps companies produce the “right product”. a trendline can be established through the points to create a velocity (work the team can complete per iteration). Of course. Quite the contrary. Agile processes like Scrum and XP use a concept called velocity. which is the amount of estimated effort a team can complete in a time-boxed iteration. Because we can change direction rapidly (every iteration) and the cost of change is low. rather than trying to guess how things might shape up a priori.21 This is largely because the unused features were specified in some up- front plan before the ratio of their cost to value was considered or even knowable. we recently delayed selecting a database package for an application because some of the desired features were not available at that time in the options we had to choose from. Once a team has established a velocity. . One of the biggest advantages to IID is that work can begin before all of the requirements are known. We therefore built the system in a database independent manner. but in the end it is the customer that decides what the development team builds. Each point of the chart in Figure 4 represents an iteration (or Sprint in Scrum). Focusing on high business value features first is therefore a critical component of efficient Agile development. Inc. This is especially the case with small businesses where domain experts wear many hats and often cannot commit to two or three months of straight requirements analysis. All rights reserved. Page 8 gives customers a chance to “try software” periodically and provide feedback. Decisions can be delayed to some future iteration when better information or technology is available to optimize the choice. Many organizations are not fully staffed with business analysts cranking out reams of requirements specs. But we already know that we cannot plan for everything in advance. The Agile answer is to examine project progress empirically. and the Y-axis represents the total estimated effort remaining for the backlog. “how can we budget for such a project?” It sounds a bit scary: let’s start working in short iterative cycles that yield demonstrable software without actually planning everything in advance. The trendline can then be Copyright © 2004 Danube Technologies. “how do you know when the software will be finished if there’s no up-front plan?” and the obligatory follow up question. As iterations progress. and (luckily) a few weeks before the product launch a new version was released by one of the database vendors that solved our problem. there is a valuable opportunity for the customer to re-examine business factors at the beginning of each iteration to select features for inclusion in the current iteration according to business ROI. Convergence with Agile One of the most commonly asked questions by those examining Agile is. How do Agile projects prioritize work? A study by the Standish Group shows that in typical software systems 45 percent of the features are never actually used by users and another 19% are only rare used.

For complex problems like project convergence. And quality is non-negotiable. To answer the questions above. the features built should always be high quality. . Inc. which represents the empirical estimate of the completion date. work will proceed on the highest priority requirements first to ensure that the most important things get done before a deadline or the money runs out. Page 9 extrapolated to determine the X-intercept. Going to production with high priority features is better than never going to production at all. There is no guarantee all features will be built.22 That is. scope. All rights reserved. but it is certain that the highest priority features will go into production and that they will be built well. perhaps the reduction in scope was accompanied by the removal of critical impediments to efficient progress. In this case. schedule. Agile methods tell us that the customer cannot specify all four of the software development variables (cost. especially considering the Standish report cited above that nearly 65% of features are never or rarely used in reality. then the scope of the work must be variable or the definition of the scope must be at a high level so the robustness of each feature can be negotiated. the project’s “burndown” trendline indicated an X-intercept well into the future (off the charts!). Figure 4: Product Burndown Chart with Velocity 1200 Velocity: 5 story pts per sprint 1000 800 Story Points 600 Velocity: 25 story pts per sprint 400 200 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Sprints Figure 3 is a Product Burndown Chart representation of a typical project. Copyright © 2004 Danube Technologies. adhering to strict code and testing standards. Notice also that the velocity (slope) changed for the better. if the timeline and cost variables are fixed. the product scope had been adjusted down so that the project could be completed by the budgeted 20th iteration. By iteration ten. Through the first nine iterations. quality).

It has the most comprehensive empirical evidence for Agile/Iterative of any book currently on the market. “Lean Software Development An Agile Toolkit”. XP. It contains the most compelling case for agile over sequential development I have yet to uncover.25 References After reading an introductory article. but how do you do agile? Scrum is a very good place to start on the management side. “Agile & Iterative Development: A Manager’s Guide”. a highly effective agile project management method. Agile is simply the latest theory that is widely replacing the waterfall approach that itself will change and evolve well into the future. Although waterfall is often referred to as “traditional”. Addison-Wesley. much of this article leans on the Poppendieck’s work.24 The notion that Agile is a radical deviation from the long established. Unlike bridge building. All rights reserved. Page 10 Conclusion But does Agile/IID work? Of course the proof is always in the pudding. 2003. Inc. Theory and rhetoric are nice. 2000. Tom Poppendieck. Copyright © 2004 Danube Technologies. software engineering has had a very short history relative to other engineering disciplines. software development is not built on thousands of years of trial and error. This book is outstanding and each page seems to offer a valuable nugget of information. 2003. In 1994. Standish reported a 31% failure rate that has improved to 15% in 2004. Kent Beck. Addison-Wesley. and the most recent 2004 Standish Group CHAOS report on the success of software projects shows a dramatic improvement in the failure rate of software projects. sometimes I wish I could ask the author for a list of good books or articles to further my knowledge on the subject. “Extreme Programming Explained: Embrace Change”. So what do your developers do differently in agile? Extreme Programming (or XP) advocates engineering principles such as pair programming and test driven development and Beck’s book is the de facto authority. and is therefore in a rapidly evolving infancy as an engineering discipline. . Ken Schwaber. Prentice Hall. “Agile Software Development with Scrum”. 23 Standish Chairman Jim Johnson attributes the improvement to smaller projects using iterative processes as opposed to the waterfall method. Craig Larman. Addison-Wesley. Below are a few good starting points for anyone interested in learning more about agile software development. This book should be read by any manager interested in Agile. It also delivers outstanding tools for implementing agile. tried and true history of waterfall software development is incorrect. and the Unified Process. 2001. It also nicely summarizes and contrasts some of the major Agile/Iterative approaches such as Scrum. Mary Poppendieck. This book is about “Scrum”. Mike Beedle.

Basili.agilemanifesto. Prentice Hall. 4 Ken Schwaber. “Chaos Chronicals”. 10 The Agile Manifesto is online at http://www. 9 I. see: Mary Poppendieck.cfm?Doc=newsletter/2004-01-15/Standish. 2000. 7 For more on how lean development influences agile software development. Computer. 15-19. “Extreme Programming Explained: Embrace Change”. “Agile Software Development with Scrum”. 2003. 137-146. Tom Poppendieck. Copyright © 2004 Danube Technologies. 2002. 12 Kent Beck. Software Magazine and Weisner Publishing. http://www1. Addison-Wesley. 1994. software developers find it very objectionable when asked to skimp on quality. Takeuchi. As professionals. 8 Craig Larman. 23 Standish Group International. 2000. Nonaka. pp. 100-101 17 In fact. Nonaka.standishgroup. year 2000. 18-19. Harvard Business Review.. pp.php 6 I. All rights reserved. Agile Modeling Essays excerpted from the book “The Object Primer. “Software Engineering Economics”.com//sample_research/chaos_1994_1. Tom Poppendieck. “ROI. September 2000. it is debatable whether Quality is really an adjustable factor.. p 28. “Is Design Dead”. period. 3rd ed. Addison-Wesley. 22 Mary Poppendieck. Mike Beedle. 2004. 2003. January 1986. XP Magazine. http://www. Published Keynote Third International Conference on Extreme Programming. 48. Software developers should always aim for high quality software. 18 Kent Beck. Inc. Takeuchi. 24 “Standish: Project Success Rates Improved Over 10 Years”. Inc. pp. Proc. 1994. Addison-Wesley. “Lean Software Development An Agile Toolkit”.. “Chaos Chronicles”. “Extreme Programming Explained: Embrace Change”.html .standishgroup. 15 Martin Fowler. 328-339. . It’s Your Job!”. 89-94 5 Standish Group 11 Barry Boehm. 2 Standish Group International. “Extreme Programming Explained: Embrace Change”. Addison-Wesley. Inc. H. “Agile Software Development with Scrum”. Mike Beedle. Addison-Wesley. 2001. p 32. 19 The Scrum Agile method has its roots in the Nonaka-Takeuchi article referenced above. “Managing the Development of Large Software Systems”. Inc. IEEE CS Press. by Scott Ambler. Westcon. “The New New Product Development Game”. pp. H.softwaremag. by Alistair Cockburn. p. Prentice Hall PTR. http://www1. 21 Jim Johnson. Harvard Business Review. Addison-Wesley. 25 Software architect Michael James has a semi-serious theory that “test-only” development will soon be feasible with the advances in cheap. Tom Poppendieck.: Agile Model-Driven Development with UML2”. clustered supercomputing.php 3 Kent Beck. 13 Reexamining the Cost of Change Curve. by Scott Ambler. January 1986. Cambridge University Press. Victor R. and for the same ethical implications software developers resent this charge. Page 11 1 Winston Royce. IEEE CS Press. Surgeons or lawyers would be sued for malpractice. Prentice Hall. 2001. “Chaos Chronicals”. 14 Examining the Cost of Change Curve. 16 Ken Schwaber. 2004. http://martinfowler. “Lean Software Development An Agile Toolkit”. “Lean Software Development An Agile Toolkit”. pp. 1981. 1970. 20 Mary Poppendieck. “Iterative and Incremental Development: A Brief History”. Developers write only robust tests and supercomputers will write and compile code until all tests pass. 2000. “The New New Product Development Game”. pp. 2003. June 2004.