The AI-DLC: The AI-driven development lifecycle

Coined by AWS’s Raja SP in 2025 and described in his “AI-Driven Development Lifecycle (AI-DLC) Method Definition” whitepaper[x], the AI-DLC methodology is presented as a machine-human collaboration, as opposed to previous methods like Agile and Scrum, which were fundamentally human-centric. The author compares this shift to the replacement of horse-drawn carriages by the automobile. But while simply adding AI to Agile produces a “faster horse”—to borrow a metaphor often misattributed to Henry Ford, the AI-DLC represents a more fundamental evolution.

Whereas AI tools have been integrated into the SDLC in recent years, the AI-DLC represents a step of further integration of AI. In short, AI is not relegated to a mere code generation tool. It’s not just a part of development, but central to it. Collaboration between AI and humans is continuous, and human engineers take on a more supervisory role.

It’s also distinct from the term “AI lifecycle,” which describes how AI models are themselves built, rather than how they are used to assist in the development of software.

Raja SP’s paper defines ten principles that represent the foundation of the AI-DLC.

AI should not simply be bolted onto existing software development methodologies. AI is not just another productivity tool—it’s a core participant. It asks what the SDLC would look like if AI had always existed. The result is an AI-centric and AI-native SDLC.

AI-assisted development traditionally looks like developers continuously prompting generative AI to perform individual tasks. Under the new paradigm, instead of a developer asking an AI for help and waiting for the AI to perform the task, the developer describes their intent, and the AI uses its agentic capabilities to create a plan, ask questions, execute, validate and maintain, all while proactively seeking approval from its human partner.

Previous methodologies let development teams choose their own design techniques. The AI-DLC centralizes design and brings engineering disciplines into the methodology itself. Practices such as Domain-driven design (DDD), Behavior-driven development (BDD) and Test-driven development (TDD) become integral to the workflow. This standardizes engineering quality.

AI can’t be 100% autonomous. The paper encourages human-in-the-loop validation and oversight of AI decisions so that AI benefits from human judgment.

In enterprise software, the complexity is about more than just writing code. Large systems are made up of multiple services, architectures, different kinds of users and stakeholders and inevitably years of tricky technical debt. One of the AI-DLC’s value propositions is that it preserves all this context across the whole project, so developers can continually benefit from this uniquely holistic understanding.

The AI-DLC revolutionizes software development, but retains the role of human participation where it’s needed. A culture of continuous feedback is maintained throughout projects to help ensure that development stays aligned with business objectives. These touchpoints are automated and documented so that humans can focus on making the best decisions.

The author recognized that organizations can’t swap in a new methodology overnight. The process is gradual and integrating with existing concepts and workflows that organizations are already familiar with, so adoption is minimally disruptive.

Since AI can perform work and maintain context across disciplines, tasks that once required separate teams of specialists can now be completed by more integrated teams with higher-level understanding. The result is less frequent handoffs.

The traditional SDLC breaks out work into distinct phases like planning, design, implementation, testing and deployment. At the end of each phase, information is passed from one to the next. The AI-DLC blurs these boundaries by treating these phases as a continuous flow where tasks occur simultaneously and non-linearly. The result is shorter feedback loops and faster progress.

The AI-DLC does not prescribe a rigid workflow for every project—it’s adaptive. It does not force projects through a predefined process. In every scenario, AI analyzes the goals and constraints and generates a customized workflow that makes the most sense for that particular project.

The framework asks the question: What if software development wasn’t built around human limitations? Previous methodologies like Waterfall and Agile were designed to help human teams coordinate their effort. AI fundamentally changes this core assumption.

Using a Spec-driven development (SDD) methodology, The AI-DLC emphasizes intents rather than rigid processes, and allows the AI to figure out how to build something that best meets those specifications. Derived from these intents are units, sets of tasks which represent the AI’s path toward accomplishing the intent. A bolt is an iteration by which units of work are executed.

Work progresses through several phases. The inception phase captures intents and translates them into units using mob elaboration, a collaborative software planning ritual where a cross-functional team and an AI system work together to translate high-level business ideas into user stories.

Requirements analysis begins here. Non-functional requirements (NFRs) define crucial attributes, guardrails and constraints. Risk assessments are conducted here as well, but both risk and requirements are continuously checked and reevaluated throughout the lifecycle.

Steering files are markdown documents stored in a repository that define the rules, constraints, architectural standards and workflows that agents must follow. They’re created during the inception phase, but are involved across all three phases of the AI-DLC. They act as a persistent source of truth for AI agents across the project.

The construction phase is where execution happens. During mob construction, the human team and AI collaboratively write, test, and deploy code together.

In the final operations phase, the AI uses the business requirements from the inception phase and the technical decisions from the construction phase to manage deployment, automate infrastructure and monitor the live system.

Across each phase of the SDLC, AI plays a central role.

In a traditional SDLC, humans manually translate business objectives into project plans. Organizations have been experimenting with workflows in which AI assistants help translate high-level goals into plans. In the AI-DLC, AI agents ask clarifying questions, identify ambiguities, make suggestions and bring up potential missed opportunities. These are converted into stories and project roadmaps. Agentic AI maintains this context throughout the project.

In the AI-DLC, AI doesn’t just summarize meetings. It analyzes emails, support tickets, transcripts and documentation to identify conflicts and gaps before development even starts. If someone brings up a tentpole event in one meeting that conflicts with the project timeline as articulated in another meeting, the agent can flag this potential discrepancy and provide a resolution. This provides an umbrella understanding of a project from start to finish that is difficult for any one individual or team to maintain.

Artificial intelligence turns software design into an iterative collaboration. AI can propose application architectures, database schemas, APIs, infrastructure configurations and user interface designs, explaining the tradeoffs between different approaches. It can quickly generate interactive prototypes, compare alternative designs and revise architectural documentation as the project evolves.

The core task of software engineering—writing code—is the most straightforward area where AI is useful in the SDLC, and where AI takes on its most autonomous role. AI coding agents understand the project context, coding standards and business logic, and can build autonomously with these in mind. Developers validate decisions and refine AI-generated code.

Testing is replaced by a broader verification phase, since AI is continuously generating and executing tests throughout development rather than waiting for it to finish. It analyzes code for defects, detects regressions, validates compliance with architectural constraints. This happens in the background at every step.

Within the AI-DLC, deployment is another continuously optimized process. Agents prepare release plans, validate infrastructure configurations, generate documentation, monitor production rollouts and optimize CI/CD pipelines. AI actively helps make sure that releases are stable and secure, as well as aligned with operational objectives.

The AI-DLC extends beyond traditional software maintenance by treating production systems as an ongoing source of feedback. By preserving knowledge across the entire lifecycle, AI enables each release to benefit from lessons learned in previous development cycles.

The paper encourages organizations to practice the various rituals and embed the AI-DLC in their own orchestration tools, with the goal of minimizing the need for significant overhauls.

The framework is standardized around AWS tools like Kiro and Amazon Q Developer. But it can also be used with agentic platforms like IBM Bob and Claude Code.

IBM Bob, for example, is an end-to-end agentic development partner that’s structurally optimized for the AI-DLC. IBM Bob features Ask and Plan modes that take high-level intents and turn build the initial application design structure. There’s Agent Mode for the writing, modifying and refactoring code or for fixing bugs or creating new files. Users can also tailor Bob’s behavior by building custom modes with specialized roles, tool restrictions and team workflows.