In the previous articles of this series we covered SDD (Spec-Driven Development) and how to apply quality gates, tests, and templates. Now I’ll show something I haven’t found clearly documented anywhere: the complete BMAD workflow, step by step, showing how AI agents organize themselves to deliver production code.
What is BMAD?#
BMAD (Breakthrough Method for Agile AI-Driven Development) is an open source framework that organizes AI agents as if they were an agile team. Each agent takes on a role, PM, Architect, Dev, QA and handles a specific part of the development workflow.
The key point is: you don’t just tell the AI “go do it”. You follow a workflow where each phase produces artifacts that feed the next, and each AI agent knows exactly what to do because it has structured context.
The Agents#
Before diving into the workflow, let’s see who does what. Remember: these are all AI agents, not real people.
| AI Agent | Role | What it does |
|---|---|---|
| Analyst | Discovery | Domain, market, and competition research |
| PM | Planning | Creates PRD with functional and non-functional requirements |
| UX Designer | Design | Creates UX specs, components, interaction patterns |
| Architect | Solutioning | Makes technical decisions (ADRs), validates readiness |
| Scrum Master | Organization | Plans sprints, creates stories, runs retrospectives |
| Dev | Implementation | Executes stories using TDD |
| QA | Quality | Test strategy, automation |
| Solo Dev | Quick Flow | Handles spec + dev + review for small changes |
Two Paths: Full Method vs. Quick Flow#
The first decision in BMAD is: how big is the change?
- Full Method: goes through all phases, with multiple AI agents collaborating
- Quick Flow: a single agent (Solo Dev) handles spec + dev + review
This separation is fundamental. Without it, you end up using the full process to fix a typo, or worse, building a complex feature without specs.
Full Method: The Complete Workflow#
Greenfield (New Project)#
When the project is new, we start from scratch, from analysis to implementation:
Phase 1 — Analysis (Analyst Agent)
The Analyst agent does the initial research: domain, market, competition, technical feasibility. It produces the Product Brief with vision, personas, and success metrics.
Phase 2 — Planning (PM and UX Designer Agents)
The PM agent creates the PRD (Product Requirements Document) with all functional requirements in Given/When/Then format. If the project has a UI, the UX Designer agent creates design specs: components, interaction patterns, design system.
Phase 3 — Solutioning (Architect Agent)
The Architect agent makes technical decisions and documents them as ADRs (Architecture Decision Records). Then it breaks the PRD into epics and stories. Finally, it runs a readiness validation to ensure everything is aligned before coding begins.
Phase 4 — Implementation (SM, Dev, and QA Agents)
This is where the code happens:
- The Scrum Master agent plans the sprint and creates story files with full context
- The Architect agent validates the Definition of Ready for each story
- The Dev agent implements using TDD: RED -> GREEN -> REFACTOR per AC
- The QA agent does adversarial code review
- Quality gates: tests + build + type-check
- Commit and artifact sync
Brownfield (Existing Code)#
When code already exists in production, the workflow changes. No Phase 1 (analysis) needed since the domain is already known. Context comes from the codebase itself:
The key difference: before anything else, you generate a project context, a document describing the existing codebase’s patterns, stack, conventions, and structure. This allows AI agents to follow existing patterns rather than inventing new ones.
Quick Flow: The Fast Path#
For small changes, bugs, config tweaks, simple features touching fewer than 3 files the Quick Flow is the way to go.
A single AI agent (the Solo Dev) handles everything:
- Quick Spec: creates a simplified tech-spec or story
- Quick Dev: implements the change
- Code Review: reviews its own code adversarially
Quick Flow doesn’t skip quality gates, tests, build, and type-check are still required. What it skips are the planning and solutioning phases that don’t make sense for small changes.
When to use Quick Flow vs. Full Method?#
| Scenario | Path |
|---|---|
| Isolated bug fix | Quick Flow |
| Config change | Quick Flow |
| Simple feature (< 3 files) | Quick Flow |
| New feature with multiple endpoints | Full Method |
| Architecture refactor | Full Method |
| New epic or module | Full Method |
| Changes affecting security/auth | Full Method |
When in doubt, use the Full Method. Better to have unnecessary specs than code without specs.
The Workflow’s Artifacts#
Each phase produces artifacts that feed the next. This is what differentiates BMAD from simply asking the AI to “just do it”:
Artifacts are living documents. When implementation reveals that something needs to change in the specs, the specs get updated. This feedback loop is what keeps everything consistent.
Sprint Status as Source of Truth#
Each story’s state is tracked in a central file:
development_status:
# Epic 1: Authentication
epic-1: in-progress
1-1-user-registration: done
1-2-user-login: in-progress
1-3-password-reset: ready-for-dev
1-4-oauth-integration: backlog
# Epic 2: Dashboard
epic-2: backlog
2-1-dashboard-layout: backlog
The possible statuses follow a linear flow:
The Complete Decision Flow#
Putting it all together, the BMAD decision flow looks like this:
Final thoughts#
BMAD isn’t a framework to make things more bureaucratic. It’s about making work with AI agents predictable and reliable. Without a structured workflow, every session is a lottery, sometimes the AI gets it right, sometimes it doesn’t.
With BMAD:
- Small changes are fast (Quick Flow)
- Large changes are well-planned (Full Method)
- All code has specs, tests, and review
- Context persists between sessions through living artifacts
If you read the other articles in this SDD series, you now have the full picture: the workflow (this article), the philosophy, and the practices.

Comments
Comments use Disqus and load only if you click the button below.