toruiwasa.com

AI-Driven Development Lifecycle (AI-DLC) – Reimagining How We Build Software

Published: 2026-02-28

AI has been quietly reshaping software development for some time now. Our productivity as engineers has improved dramatically — tasks that once took days are now drafted in minutes. Beyond that, a new phenomenon called Vibe Coding has started to democratise software creation: today, anyone can build working software without a traditional programming background.

Against that backdrop, a post published on the AWS Blog on July 31, 2025 attracted industry attention. The title: AI-Driven Development Life Cycle: Reimagining Software Engineering.

What made it stand out was its premise. Rather than shoehorning AI into the rituals of existing SDLC methods, it proposed building a new framework from the ground up — one where AI is assumed to be present at every stage, not just as a coding assistant. AI handles planning, decomposition, domain modelling, code generation, and testing, while Product Owners and developers remain indispensable as supervisors and decision-makers who validate what AI produces and steer it toward the right outcomes.

Admittedly, much of what the AI-DLC white paper describes is already achievable — in spirit — with agentic tools like Claude Code today. But AI-DLC reads more like the enterprise-grade, structured version of that experience: explicit phases, defined artifacts, mandatory human checkpoints. And perhaps more importantly — have you seen another source that has articulated an AI-centric development approach at this level of detail? I hadn’t.

That alone made it worth studying closely.

And on a personal note — the more I read it, the more I became convinced that AI-DLC isn’t just for large engineering teams. The methodology’s core loop (AI proposes → human validates → AI executes) works equally well when the “team” is just you. I’ll come back to that.


Why AI-DLC? The Problem With Retrofitting

Before understanding what AI-DLC proposes, it’s worth internalising what it’s reacting against.

In the AI-Assisted era — where most teams operate today — AI augments individual tasks: write this function, generate this test, summarise this PR. Developers still do the heavy lifting. AI is a very smart autocomplete.

The problem is structural. Existing SDLC methods enforce long iteration cycles, rigid role specialisation silos, and manual planning overhead — standups, retrospectives, estimation sessions — that AI could handle in seconds. In this model, AI sits on the periphery, assisting discrete tasks rather than orchestrating the workflow.

Retrofitting AI into these constraints doesn’t unlock its potential. It just makes old processes marginally faster while preserving their fundamental inefficiencies.

AI-DLC’s answer is first-principles thinking. Rather than asking “how do we use AI in Scrum?”, it asks: “what would software development look like if we designed it specifically for AI, from scratch?”


What is AI-DLC? The 10 Core Principles

AI-DLC is built on ten design principles. Rather than listing them abstractly, here’s what each means in practice:

1. Reimagine Rather Than Retrofit

Don’t patch Agile. Build a new mental model where AI is the engine, not the passenger.

2. Reverse the Conversation Direction

AI initiates and directs. You validate and decide. Think of it like Google Maps: you set the destination (the intent), the system handles step-by-step routing, and you course-correct when needed. The direction of the conversation flips — and that changes everything.

3. Integrate Design Techniques Into the Core

AI-DLC embeds design methodologies — like DDD, TDD, and BDD — into its core process, rather than treating them as optional add-ons. AI applies them during planning and decomposition, so you always work within coherent, well-structured boundaries instead of retrofitting best practices after the fact.

4. Align With Current AI Capability

AI-DLC is optimistic but realistic. It doesn’t assume AGI. Current AI can plan, generate, and reason well — but still needs human oversight for contextual judgment. The methodology is designed to work now, not in some hypothetical future.

5. Cater to Complex Systems

AI-DLC targets systems that demand continuous adaptability, high architectural complexity, and multi-team coordination — the kind with numerous trade-offs, integration requirements, and regulatory constraints. Simpler systems that need few trade-offs are better served by low-code or no-code approaches.

6. Retain What Enhances Human Symbiosis

Not everything gets thrown out. User stories remain — they’re a valuable alignment contract between humans and AI. Risk registers stay. Acceptance criteria stay. What AI-DLC drops is the ceremony that no longer serves a purpose when AI handles the underlying work.

7. Facilitate Transition Through Familiarity

Familiar concepts get reimagined, not discarded. Sprints become Bolts — shorter, more intense cycles measured in hours or days rather than weeks. Epics become Units — cohesive, independently buildable work elements. The shift in vocabulary signals the shift in pace.

8. Streamline Responsibilities for Efficiency

As AI assumes responsibility for cross-cutting execution — security patterns, infrastructure decisions, and test generation — developers can remain engaged across the full stack as supervisors and decision-makers.

Rather than eliminating expertise, this model reduces fragmentation of responsibility. By absorbing cognitive load, AI enables teams to avoid deep role segmentation, minimising knowledge silos and context loss caused by handoffs.

The result is not fewer humans, but fewer artificial boundaries — leading to clearer accountability, leaner team structures, and more coherent system ownership.

9. Minimise Stages, Maximise Flow

Reducing handoffs means increasing flow. AI-DLC compresses weeks of sequential planning into hours via rituals like Mob Elaboration, with each phase’s output serving as direct input for the next. Human validation acts as a checkpoint — a loss function — that catches errors before they compound downstream.

10. No Hard-Wired Workflows

There’s no prescriptive “green-field workflow” versus “bug fix workflow”. Instead, AI interprets the intent and proposes a Level 1 Plan, which humans iteratively refine through dialogue.

Execution remains AI-driven, while humans retain oversight and validation authority. The methodology adapts to the nature of the work — and evolves alongside AI capabilities — rather than forcing work into predefined procedural moulds.


The Three Phases

AI-DLC operates in three phases, each producing artifacts that serve as “context memory” for the next. Every decision, every validated output, gets persisted — giving the AI increasingly rich context as the work progresses.

🔵 Inception Phase — Translating Intent into Execution-Ready Units

The Inception Phase converts a high-level business intent into loosely coupled, measurable Units designed for parallel and autonomous execution downstream.

The key ritual is Mob Elaboration: the entire team — Product Owner, developers, QA, stakeholders — sits together with a shared screen. AI begins by asking clarifying questions to eliminate ambiguity and fully understand the business goal. It then proposes user stories, non-functional requirements, and risk descriptions, and composes cohesive stories into Units. The team reviews, refines, and corrects under- or over-engineered parts, aligning them with real-world constraints. The Product Owner validates and adjusts where necessary.

What used to take weeks of meetings condenses into hours — and the output isn’t just a backlog. It’s a structural execution design: a set of independently deployable units, each with clear scope and measurable value.

Artifacts produced:

  • User Stories with acceptance criteria
  • NFR definitions (performance, security, scalability)
  • Risk descriptions mapped to the org’s risk register
  • Units — cohesive, loosely coupled functional blocks, analogous to Subdomains in DDD or Epics in Scrum
  • Measurement Criteria — tracing each Unit back to the original business intent
  • Suggested Bolts — the iteration cycles that will execute the Units
  • PRFAQ (optional) — a press release and FAQ that crystallises the business intent

🟢 Construction Phase — Determine How to Build It

This is where the structural design from Inception turns into working software. Execution happens within the Bolts (iteration cycles) proposed earlier, allowing teams to deliver different Units in parallel.

For each Unit, development progresses through four steps:

  1. Domain Design — AI models the core business logic using DDD: entities, value objects, aggregates, domain events, repositories, and factories — independently of infrastructure concerns. You validate and refine.
  2. Logical Design — The domain model is extended with non-functional requirements. AI recommends architectural patterns (CQRS, Circuit Breakers, Event-Driven, etc.) and generates Architecture Decision Records (ADRs) for your review.
  3. Code & Test Generation — AI writes the executable code mapped to appropriate services and drafts the corresponding test suites (functional, security, performance). Crucially, developers review and adjust both the code and the test scenarios before any execution happens.
  4. Execution & Validation — AI runs the approved tests, analyses any failures, and proposes fixes. Developers validate the findings and approve the resolutions.

For brown-field scenarios (existing systems), Construction begins differently — AI first elevates the codebase into higher-level models: static models of components, responsibilities, and relationships, and dynamic models showing how components interact for the most significant use cases. With that context established, the standard steps follow.

The ritual for this phase is Mob Construction. Because Units are often being built concurrently, teams use this collaborative time to exchange integration specifications — like APIs or event schemas — defined during the design steps. AI handles the heavy lifting of the implementation, while developers focus on managing these cross-team boundaries, making trade-off decisions, and validating the output.

🟡 Operations Phase — Deploy and Monitor

The Operations Phase handles packaging, deployment, and ongoing monitoring.

AI packages each Unit into Deployment Units — container images, serverless functions, IaC stacks (Terraform, CDK, CloudFormation) — generating and executing a comprehensive test suite covering functional acceptance, security, and load testing. In production, AI continuously analyzes telemetry, predicts potential SLA violations, and proposes concrete mitigations: scaling a service, rebalancing traffic, increasing throughput. Developers review and approve before any action is taken.


Key Artifacts At a Glance

ArtifactWhat It Is
IntentHigh-level goal that starts the workflow
UnitA self-contained, loosely coupled work element (replaces Epic)
BoltShortest iteration cycle, hours to days (replaces Sprint)
Domain DesignDDD model of business logic — entities, aggregates, events
Logical DesignDomain Design + NFRs + architectural patterns + ADRs
Deployment UnitPackaged, tested, operations-ready output — container/serverless + IaC

AI-DLC for Personal Projects

Here’s something the methodology paper doesn’t say out loud: you don’t need a mob for Mob Elaboration.

The rituals are designed for enterprise teams, but the underlying principle is universal. When you state an intent and let AI ask clarifying questions before writing a single line of code, you’re doing Mob Elaboration solo. When AI generates a domain model and asks you to validate it before generating code, you’re doing Mob Construction.

For a solo developer, AI-DLC shifts the dynamic from:

“I think I know what I want to build. Let me start coding and figure it out.”

To:

“I have an intent. Let me let AI break it down, surface what I’ve missed, and identify the decisions I actually need to make.”

This isn’t about adding ceremony. It’s about shifting from reactive to intentional development. AI-DLC gives you a structured way to have a conversation with yourself — mediated by AI — before you start building. The result: less rework, clearer scope, and better-quality output.

I’ve been using these principles on this very blog’s codebase. The difference is real.


How to Adopt AI-DLC

AI-DLC doesn’t replace or reject existing agile practices; rather, it extends them as an evolved practice model.

It is designed as a reproducible set of rituals, such as Mob Elaboration and Mob Construction, which become internalized through repeated practice. The AI-DLC Unicorn Gym by AWS serves as an accelerator to support practical implementation.

Additionally, for large-scale organizations, AI-DLC can be embedded into existing workflow tools, enabling seamless adoption without the need for major change initiatives.


Wrap Up

AI-DLC is not just an incremental improvement – it represents a fundamental shift in human–AI interaction.

AI is already capable of taking ownership of planning, decomposition, modelling, and even code generation. The highest-leverage contribution a developer can make today is not writing more code, but making better decisions about what gets built and how.

AI-DLC serves as a practical guide for developers who have not yet fully unlocked the power of AI.

Whether you’re leading a large engineering organization or building a side project on weekends, that shift is available to you right now.


Resources


Keep on building.

  • Blog