Why Gen AI Won’t Kill Agile - It’ll Supercharge it

Current Gen AI coding tools benefit from a large amount of well-structured context describing all aspects of the problem before a single line of code is generated. Simply trying to 'one-shot' a solution is next to impossible, no matter how good your prompt-writing skills are.

The approach followed by many has been to have an AI agent (or set of agents) assist in a detailed problem definition exercise when starting out. The human in the interaction can provide a rough summary of the problem to be solved and ask the AI what clarifications it needs in order to proceed - in much the same way a Product Owner and Tech Lead might interact. From there, artifacts such as Problem Description, Requirements Documents, Architecture Diagrams, Use Cases, and even a Project Development Plan are created (automatically) and persisted as a snapshot of all necessary project information. Only once all this documentation has been completed should any coding happen. The teams at our AI Cookoff that followed this approach had the most success in completing their chosen applications within the timeframe with the least amount of headache.

What Does This Mean for Agile Development?

The process described above has all the hallmarks of Waterfall SDLC - lots of up-front design, extensive documentation, and the implicit assumptions that what gets documented is what will be built, and what is documented is what the customer actually wants. This is a very different approach to Agile, which is all about iterative development and delivering value to the customer in small increments.

At Electric Mind, Agile development and delivery has been part of our DNA for 30 years. One of our core beliefs is that through frequent iteration, evolutionary design, and a flexible, adaptable mindset we can deliver better systems faster and with lower risk than through traditional Waterfall approaches. Does this mean that the Gen AI tools are not a good fit for Agile development? Or more likely, is our chosen methodology incompatible with the trajectory software development is taking into the future? Our instinct says 'no' and the reason comes down to opinions about the core values of Agile.

The Core Values of Agile

Modern Agile development comes with a wide array of tools, practices, products, and reports designed to help teams and stakeholders tackle the challenges of delivering software in a fast-paced, ever-changing world. It can be very easy to lose sight of the simplicity of the core value of Agile with the complexity of the tools and practices. In our view, the two most important values of Agile are:

1. Embrace change as a natural, unavoidable part of the development process. Make it part of your mindset.
2. Deliver working software in small increments. Get something working, then keep it working as you iterate on it.

Although the Gen AI coding tools seem to have the most success when using a documentation-first approach, this documentation is not used for the same purpose as traditional Waterfall development - that is to reduce or eliminate the possibility of requirements changing during the development process. Instead, the documentation created when interacting with Generative AI models is merely the interface to a fast-paced tool which enables extremely quick iteration. In fact, it is often the case that the most effective way to make a code change is to tweak the requirements document, architecture diagram, or other design artifacts and allow the tooling to update the codebase as required. In this way Gen AI coding tools support the flexibility and speed of change required for Agile development. It is still up to the developer and stakeholder to have the correct mindset.

Working Software in Small Increments

To support the second core value, Gen AI tools excel at standing up running applications from scratch. It is relatively simple to get a new 3-tier application running on your developer machine in a small number of prompts. Keeping it running, however, is a larger challenge. It is quite common for successive changes to break a running system, as the AI begins to hallucinate or make compounding errors which result in brittle or broken software. The solution to this is not to attempt to completely specify our application's behaviour up front and hope for a big bang integration. Instead, we have found success in trying to _limit_ the scope of each AI-generated change - keep it on a leash, make small incremental or iterative changes, and avoid large, sweeping refactorings. Protect the system against regression by creating a robust suite of automated tests, and by reviewing each change before merging it to main branches (hint: these activities can be performed by AI tools as well).  This, of course, is standard practice on any software development project, regardless of whether the code was written by a human or a machine.

A Golden Age for Agile Software Development

In short, generative AI coding tools do not spell the end for Agile development. Quite the opposite in fact - we believe that we are entering a golden age for Agile software development. Generative AI coding tools will enable iteration, experimentation, and a rate of change unseen before now. Iterations will be measured in minutes and hours, not days or sprints, all while maintaining the mindset and mechanisms of Agile development. With this boost in productivity comes the freedom to experiment, and the confidence to invest extra effort in quality.

Wondering how to make Gen AI work inside your Agile approach? We’ve tested the approaches, seen what works (and what doesn’t), and know how to help teams get started. We’re here to help turn your ideas into software that works and works fast.

‍