AI-DLC 2026：人機協同，重塑自主AI的開發模式

AI-Driven Development Lifecycle 2026 (AI-DLC 2026)

A Methodology for the Age of Autonomous Agents

A comprehensive methodology reimagining software development for the era of autonomous AI agents, introducing human-on-the-loop workflows, backpressure-driven quality, and the Ralph Wiggum autonomous loop pattern.

Acknowledgments & Attribution

This methodology synthesizes foundational work from the AI development community with lessons learned from production deployments of autonomous AI systems.

Foundational Work

Raja SP, Amazon Web Services — AI-Driven Development Lifecycle (AI-DLC) Method Definition (July 2025). The core concepts of Intent, Unit, Bolt, Mob Elaboration, and the philosophy of reimagining development methods rather than retrofitting AI into existing processes originate from this foundational work.

Key Influences for 2026

Geoffrey Huntley — Creator of the Ralph Wiggum Software Development Technique. The philosophy of "deterministically bad in an undeterministic world" and autonomous loop patterns are central to AI-DLC 2026.

Boris Cherny & Anthropic — Ralph Wiggum plugin for Claude Code, demonstrating production viability of autonomous development loops.

Steve Wilson (OWASP) — Human-on-the-Loop governance frameworks and the articulation of HITL vs HOTL operating modes.

paddo.dev — Analysis of SDLC collapse, the "19-agent trap," and the insight that phase gates become friction rather than quality control in AI-driven workflows.

HumanLayer — 12 Factor Agents principles and context engineering research including the "dumb zone" phenomenon.

Preface: The State of AI-Driven Development

Software development has undergone a fundamental transformation. What began as AI assistance for fine-grained tasks—code completion, bug detection, test generation—has evolved into AI autonomy for sustained, multi-hour reasoning and implementation. This shift demands new methodologies built from first principles rather than adaptations of existing processes.

The landscape of AI-driven development in 2026 differs dramatically from just two years prior:

AI-DLC 2026 preserves foundational insights while incorporating lessons from production deployments: the Ralph Wiggum autonomous loop methodology, the emerging consensus that traditional SDLC phases are collapsing, the practical realization that simpler AI workflows outperform elaborate orchestrations, and the understanding that organizational knowledge—previously locked in tickets, documents, and runbooks—can now serve as memory for AI agents.

I. Context

The Evolution of Software Engineering

The evolution of software engineering has been a continuous quest to enable developers to focus on solving complex problems by abstracting away lower-level, undifferentiated tasks. From early machine code to high-level programming languages, from the adoption of APIs and libraries to cloud services, each step has significantly boosted developer productivity by moving humans further from implementation details and closer to problem expression.

The integration of Large Language Models marked a revolutionary shift, introducing conversational natural language interactions for tasks like code generation, bug detection, and test creation. This was the AI-Assisted era—AI enhancing fine-grained, specific tasks while humans retained full control of workflow and decisions.

We have now entered the AI-Autonomous era. Models capable of multi-hour sustained reasoning, combined with tools for autonomous development loops, enable workflows where humans define destinations and guardrails, then step back while AI iterates toward success. Independent evaluations estimate that frontier models can now complete tasks that take humans four to five hours. Anthropic's Claude Code lead reported writing 40,000 lines of production code using Claude Code itself in a single month.

The Problem with Traditional Methods

Existing software development methods—Waterfall, Agile, Scrum—were designed for human-driven processes with long iteration cycles. Their reliance on manual workflows and rigid role definitions limits the ability to fully leverage AI capabilities. Retrofitting AI into these methods not only constrains its potential but reinforces outdated inefficiencies.

Traditional phase boundaries—requirements → design → implementation → testing → deployment—existed because iteration was expensive. When changing requirements meant weeks of rework, sequential phases with approval gates made economic sense. Each phase required:

With AI, iteration costs approach zero. You try something, it fails, you adjust, you try again—all in seconds, not weeks. The phases aren't just being augmented; they're collapsing into continuous flow.

"Phase gates that once provided quality control now create friction."

The New Reality

To fully leverage AI's transformative power, development methods need reimagination. This reimagination requires AI to be a central collaborator, with workflows, roles, and iterations aligned to enable faster decision-making, seamless task execution, and continuous adaptability.

This paper introduces AI-DLC 2026, a methodology that embraces both supervised and autonomous modes of AI collaboration, preserves human judgment where it matters most, and sets foundations for the next evolution in software engineering.

II. Core Principles

The following principles form the foundation of AI-DLC 2026. They shape its phases, roles, artifacts, and rituals. These principles are critical for validating the methodology, as they provide the underpinning rationale behind its design.

1. Reimagine Rather Than Retrofit

We choose to reimagine a development method rather than keeping existing methods like Waterfall or Agile and retrofitting AI into them. Traditional methods were built for longer iteration durations—weeks and months—which led to rituals like daily standups, sprint planning, and retrospectives. These rituals assume a cadence that AI has rendered obsolete.

Proper application of AI leads to rapid cycles measured in hours or even minutes. This demands continuous, real-time validation and feedback mechanisms, rendering many traditional rituals less relevant:

Would effort estimation (story points) be as critical if AI diminishes the boundaries between simple, medium, and hard tasks? When an AI can implement a feature in minutes regardless of apparent complexity, human estimation becomes unreliable.

Would velocity metrics be relevant, or should we replace them with business value delivered? When the constraint shifts from implementation speed to requirement clarity, traditional productivity metrics miss the point.

Would sequential phases help when try-fail-adjust cycles take seconds? When iteration is nearly free, upfront design becomes a tax rather than an investment.

These new dynamics warrant reimagination based on first principles thinking rather than retrofitting. We need automobiles, not faster horse chariots.

2. Human-on-the-Loop, Not Just Human-in-the-Loop

AI-DLC 2026 introduces the concept of AI initiating and directing conversations, using the Google Maps analogy: humans set the destination, AI provides step-by-step directions, humans maintain oversight. This methodology distinguishes two distinct operating modes:

Human-in-the-Loop (HITL): Human judgment is directly involved in decision-making. AI proposes options, human validates, AI executes. The human approves each significant step before proceeding.

This mode is essential for:

Human-on-the-Loop (HOTL): The system operates autonomously while humans monitor and intervene when needed. AI executes within defined boundaries until success criteria are met, alerting humans only when intervention is required.

This mode is appropriate for:

The Google Maps analogy extends: in HITL mode, you tell the GPS each turn to make, it confirms, you approve, it executes. In HOTL mode, you set the destination, define acceptable routes (no toll roads, avoid highways), and the navigation system handles the journey—alerting you only for unexpected detours or when intervention is required.

The key insight: The human doesn't disappear. The human's function changes—from micromanaging execution to defining outcomes and building quality gates.

3. Backpressure Over Prescription

Traditional methodologies prescribe how work should be done. Detailed process steps, code review checklists, and implementation patterns create rigid workflows that constrain AI's ability to leverage its full capabilities.

AI-DLC 2026 introduces a different approach: backpressure—quality gates that reject non-conforming work without dictating approach.

"Don't prescribe how; create gates that reject bad work."
— Geoffrey Huntley

Instead of specifying "first write the interface, then implement the class, then write unit tests, then integration tests," define the constraints that must be satisfied:

Let AI determine how to satisfy these constraints. This approach offers multiple benefits:

The philosophy can be summarized as: "Better to fail predictably than succeed unpredictably." Each failure is data. Each iteration refines the approach. The skill shifts from directing AI step-by-step to writing criteria and tests that converge toward correct solutions.

4. Embrace the Collapsing SDLC

Traditional SDLC phases existed because iteration was expensive. Each handoff between analyst → architect → developer → tester → operations lost context, added latency, and created opportunities for misalignment. Sequential phases with approval gates were an economic optimization for a world of expensive iteration.

With AI, that economic calculus inverts. Iteration is nearly free. Context loss from handoffs becomes the dominant cost. AI-DLC 2026 models development as continuous flow with strategic checkpoints rather than discrete phases.

Work stops completely at each handoff. Context transfers between specialized roles. Each new party must rebuild understanding.

Work pauses briefly at checkpoints. Same agent continues with feedback. Context is preserved throughout.

Checkpoints differ from handoffs:

This doesn't mean structure disappears. It means structure changes form—from sequential gates to parallel loops with human oversight at strategic moments.

5. Context Is Abundant—Use It Wisely

Modern language models offer context windows ranging from 200K tokens (Claude Opus 4.5) to over 1 million tokens (Claude Sonnet 4.5, Gemini). This abundance fundamentally changes how we think about AI workflows—but not in the ways that might be obvious.

The 19-Agent Trap: Early enthusiasm for AI led to complex multi-agent scaffolding that mapped AI agents to traditional org charts—an Analyst agent, PM agent, Architect agent, Developer agent, QA agent, and so on. This approach consistently performs worse than simpler alternatives because:

"As agents accumulate tools, they get dumber."

Research shows model performance degrades when context windows exceed 40-60% utilization—the "dumb zone." Information "lost in the middle" of large contexts receives less attention than information at the beginning or end.

The Insight: Abundant context windows don't mean we should fill them with everything. They mean we can be selective about high-quality context rather than compressed summaries. AI-DLC 2026 favors:

The best workflows aren't complex orchestrations—they're simple loops with clear objectives and rich, relevant context.

6. Memory Providers Expand Knowledge

AI context windows reset between sessions. Modified files and git history provide persistence without complex memory infrastructure. AI-DLC 2026 extends this insight by recognizing that existing organizational artifacts are memory providers that AI agents can access.

Traditional SDLC processes created extensive documentation that often went unused:

These artifacts represent institutional memory—decisions made, rationales documented, patterns established. Through modern integration protocols (MCP servers, API connectors, knowledge bases), AI agents can now access this memory directly.

Memory Layers in AI-DLC 2026:

Practical implications:

The filesystem remains the simplest, most robust memory provider. Git history shows what was attempted. Modified files persist across iterations. But these aren't the only options—organizational memory, properly connected, dramatically expands what agents can know.

7. Completion Criteria Enable Autonomy

The key enabler of autonomous operation is programmatic verification. If success can be measured by machines, AI can iterate toward it without human intervention for each step.

Every work element in AI-DLC 2026 should have explicit Completion Criteria—programmatically verifiable conditions that define success:

Good completion criteria are:

Examples:

This transforms the human role from "validator of each step" to "definer of done." The human specifies what success looks like; the AI figures out how to achieve it.

8. Design Techniques Are Tools, Not Requirements

AI-DLC 2026 takes a pragmatic approach: design techniques like Domain-Driven Design (DDD), Test-Driven Development (TDD), and Behavior-Driven Development (BDD) are valuable tools to apply when appropriate, not mandatory steps in every workflow.

Use structured design (DDD, TDD, BDD) when:

Skip heavyweight upfront design when:

AI can apply design patterns during execution without requiring explicit design phases. When an AI generates code, it draws on its training to apply appropriate patterns—factory methods, dependency injection, repository patterns—without being told to do so.

The test suite, not the architecture document, becomes the source of truth. If the tests pass and the code meets non-functional requirements, the implementation is valid regardless of whether it matches a pre-specified design.

9. Streamline Responsibilities

AI's ability to perform task decomposition, code generation, testing, documentation, and deployment reduces the need for specialized roles. A single developer supervising AI can accomplish what previously required separate specialists for frontend, backend, infrastructure, testing, and documentation.

This doesn't eliminate human value—it concentrates it on activities where human judgment is essential:

The role shifts from "doing the work" to "defining what work matters and verifying it's done well."

However, humans remain integral. Product owners ensure alignment with business objectives. Developers maintain design quality and handle judgment calls. These roles ensure that automation and human accountability remain balanced.

10. Platform Agnostic

AI-DLC 2026 is intentionally cloud-agnostic and platform-agnostic. The methodology applies regardless of infrastructure choices:

The methodology should be independent of vendor choices. Choose infrastructure based on requirements, cost constraints, and team expertise—not methodology constraints.

III. Core Framework

This section outlines the core framework of AI-DLC 2026, detailing its artifacts, phases, rituals, and workflows.

1. Artifacts

An Intent is a high-level statement of purpose that encapsulates what needs to be achieved, whether a business goal, a feature, or a technical outcome (e.g., performance scaling, security improvement). It serves as the starting point for AI-driven decomposition into actionable tasks, aligning human objectives with AI-generated plans.

Every Intent includes explicit Completion Criteria—programmatically verifiable conditions that define success and enable autonomous execution.

Example Intent:

A Unit represents a cohesive, self-contained work element derived from an Intent, specifically designed to deliver measurable value. Units are analogous to Bounded Contexts in Domain-Driven Design or Epics in Scrum.

Characteristics of well-defined Units:

Each Unit encompasses:

The process of decomposing Intents into Units is driven by AI, with developers and Product Owners validating and refining the resulting Units to ensure alignment with business and technical objectives.

Units should be sized for parallel autonomous execution. If a Unit requires constant coordination with other Units, it's too tightly coupled and should be restructured.

A Bolt is the smallest iteration cycle in AI-DLC 2026, designed for rapid implementation of a Unit or a set of tasks within a Unit. The term "Bolt" (analogous to Sprints in Scrum) emphasizes intense focus and high-velocity delivery, with build-validation cycles measured in hours or days rather than weeks.

Bolts operate in two modes:

Human validates each major step before proceeding. AI proposes, human reviews, AI implements, human validates. Used for judgment-heavy, high-risk, or novel work.

AI iterates until completion criteria are met, using test results and quality gates as feedback. Human reviews final output. Used for well-defined tasks with programmatic verification.

Autonomous Bolt characteristics:

A Unit may be executed through one or more Bolts, which may run in parallel or sequentially. AI plans the Bolts; developers and Product Owners validate the plan.

Completion Criteria are explicit, programmatically verifiable conditions that define when work is done. They enable autonomous execution by giving AI clear targets to iterate toward.

Completion Criteria should be:

Completion Criteria can include:

Deployment Units are operational artifacts encompassing everything needed to run in production:

AI generates all associated tests, including:

After human validation of test scenarios and cases, AI executes the test suites, analyzes results, and correlates failures with code changes, configurations, or dependencies. Deployment Units are tested for functional acceptance, security compliance, NFR adherence, and operational risk mitigation, ensuring readiness for seamless deployment.

Deployment Units should be independently deployable and include automated rollback procedures.

2. Phases & Rituals

AI-DLC 2026 organizes work into three phases, each with distinct rituals and human-AI interaction patterns.

The Inception Phase focuses on capturing Intents and translating them into Units with clear Completion Criteria for development.

Mob Elaboration Ritual

The central ritual of Inception is Mob Elaboration—a collaborative requirements elaboration and decomposition session. This happens with stakeholders and AI working together, either in a shared room with a screen or via collaborative tools.

During Mob Elaboration:

Mob Elaboration condenses weeks of sequential work into hours while achieving deep alignment between stakeholders and AI.

Outputs of Inception:

The Construction Phase transforms Units into tested, deployment-ready artifacts through Bolts. This phase progresses through domain modeling, logical design, code generation, and testing—though these steps may be implicit rather than explicit depending on complexity.

Mode Selection

The first decision in Construction is mode selection for each Bolt:

Supervised Construction (HITL)

For novel domains, architectural decisions, or high-judgment work:

Autonomous Construction (HOTL)

For well-defined tasks with programmatic verification:

Human review of autonomous work:

Mob Construction Ritual

For complex systems requiring multiple Units built in parallel:

The Operations Phase centers on deployment, observability, and maintenance of systems, leveraging AI for operational efficiency.

Deployment:

Observability and Monitoring:

AI actively analyzes telemetry data:

Anomaly Response:

For well-defined operational scenarios with runbook coverage, AI operates autonomously (HOTL for operations). Humans monitor dashboards, receive alerts for novel situations, and intervene for edge cases outside runbook coverage.

Autonomous Operations Boundaries Example:

3. The Workflow

The complete AI-DLC 2026 workflow integrates all phases into a continuous flow:

Key principles of the workflow:

IV. AI-DLC 2026 in Action: Greenfield Development

Scenario: Product Owner initiates development of a recommendation engine for cross-selling products.

1. Inception Phase (Mob Elaboration)

Human provides Intent:

AI asks clarifying questions:

"I'll help elaborate this Intent. Let me ask some clarifying questions:

Team provides answers, AI elaborates into Units:

Unit 1: Data Collection & Processing
Capture user behavior (views, clicks, purchases), build feature pipelines, create user and product embeddings.

Completion Criteria:

Unit 2: Recommendation Model
Train and serve ML model for product recommendations.

Completion Criteria:

Unit 3: API Integration
Real-time serving API integrated with product catalog.

Completion Criteria:

Unit 4: Frontend Integration
Display recommendations in product pages and cart.

Completion Criteria:

AI recommends Bolt modes:

2. Construction Phase

Unit 3 (API Integration) assigned for construction as Autonomous Bolt.

Rationale for autonomous mode:

AI execution trace:

Human reviews autonomous output:

3. Operations Phase

Deployment:

Autonomous monitoring configured:

V. AI-DLC 2026 in Action: Brownfield Development

Scenario: Adding a new wishlist feature to an existing e-commerce application.

Key Difference: Context Building

Before Inception, AI must understand the existing codebase. This analysis can itself be an Autonomous Bolt:

Autonomous context building:

Static Model captures: components, descriptions, responsibilities, relationships
Dynamic Model captures: how components interact for significant use cases

After context building, Inception proceeds normally but with awareness of existing patterns and constraints. The AI's proposals align with established conventions, and completion criteria reference existing test patterns.

Pre-Inception Artifacts

AI can also query organizational memory during brownfield work:

This context informs the Inception phase, ensuring new work aligns with existing decisions and avoids repeating past mistakes.

VI. Decision Framework: Supervised vs. Autonomous

Decision Tree

For any new task, follow this decision process:

Does the task have clear completion criteria?

Can success be verified programmatically?

Is this a high-risk change?

Is this a novel domain?

Quick Reference

Transitioning Between Modes

Work can transition between modes as understanding develops:

The flexibility to switch modes mid-work is a key feature of AI-DLC 2026.

VII. Implementing Autonomous Bolts

The Ralph Wiggum Pattern

The Ralph Wiggum pattern, named after the Simpsons character, embraces the philosophy of "deterministically bad in an undeterministic world." Rather than trying to be perfect, the agent tries, fails, learns from failures, and iterates until success.

Core components:

Autonomous Bolt Template:

Backpressure Configuration

Quality gates are the primary mechanism for implementing backpressure in AI-DLC 2026. Rather than prescribing how AI should work, gates define what must be satisfied—creating rejection signals that guide autonomous iteration toward correct solutions:

Note: Modern AI development platforms can automate quality gate configuration through plugin systems and declarative hooks, detecting project tooling (test frameworks, linters, type checkers) and configuring appropriate gates without manual YAML authoring.

File-Based Memory

Progress persists in files, enabling resumption across sessions:

Safety Limits

Autonomous execution requires safety boundaries:

VIII. Adoption Path

For Teams Already Using AI-Assisted Development

Weeks 1-2: Foundation

Weeks 3-4: First Autonomous Work

Month 2: Scaling

For Teams New to AI-Driven Development

Start with Mob Elaboration. The ritual provides:

Then gradually introduce Construction phase patterns:

Organizational Considerations

Governance:

Skills Evolution:

Metrics Shift:

IX. Appendix: Prompt Patterns

Setup Prompt

Inception: Mob Elaboration

Construction: Supervised Bolt

Construction: Autonomous Bolt

Operations: Incident Analysis

X. Glossary

XI. References

Raja SP. AI-Driven Development Lifecycle (AI-DLC) Method Definition. Amazon Web Services, July 2025. https://aws.amazon.com/blogs/devops/ai-driven-development-life-cycle/

Geoffrey Huntley. Ralph Wiggum Software Development Technique. 2025. https://ghuntley.com/ralph/

Anthropic. Ralph Wiggum Plugin for Claude Code. 2025.

Steve Wilson. Human-on-the-Loop: The New AI Control Model That Actually Works. The New Stack, August 2025.

paddo.dev. The SDLC Is Collapsing. 2025.

paddo.dev. The 19-Agent Trap. January 2026.

HumanLayer. 12 Factor Agents. 2025.

Anthropic. Claude Code: Best Practices for Agentic Coding. 2025.

Model Context Protocol (MCP). Specification and Server Implementations. 2025.

Karpathy, Andrej. LLMs in Software Development. 2025.

AI-DLC 2026 is an open methodology. Contributions and adaptations are welcome.

View source on GitHub →

Han

Built by The Bushido Collective - developers committed to honor, quality, and continuous improvement.

Links

Philosophy

"Beginning is easy - continuing is hard."

• Japanese Proverb

MIT License - Walk the way of Bushido. Practice with Discipline. Build with Honor.