BMAD-METHOD/docs/learn/module-14-agentic-development/module-14-agentic-developme...

8.4 KiB

Module 14: Agentic Development

Time: 60 min | Agents: Freya + Idunn | Phase: Design-Build


What Is Agentic Development?

Agentic development is every process where you use an AI agent to transform your ideas into coded form.

It can produce:

  • Concepts — quick explorations to test an idea
  • Proof of concepts — does this approach work at all?
  • Prototypes — working versions to evaluate and refine
  • Production code — final implementation for delivery

It works in both directions:

  • Inspiration → Specification — Ask the agent to suggest or dream up a component, section, page, scenario, or complete app. Use the output as input for your specifications.
  • Specification → Code — Feed the agent your specifications and generate working code that matches.

Whether you're exploring possibilities or building the final product, the process is the same.


The 7-Activity Menu

Agentic development is menu-driven, not linear. When you enter Phase 5, you choose what you need:

[P] Prototyping           — Build throwaway prototypes to validate ideas
[D] Development           — Write production code from approved specs
[F] Bugfixing             — Systematically fix bugs in existing code
[E] Evolution             — Add features to existing products
[A] Analysis              — Investigate codebase, architecture, patterns
[R] Reverse Engineering   — Capture existing product as WDS specifications
[T] Acceptance Testing    — Verify implementation against specs

Each activity has its own workflow with dedicated steps. The core agentic loop (plan → execute → evaluate → update) applies to all of them — the context differs.


The Agent Dialog

Every agentic development session starts with an Agent Dialog — a structured document that organizes the work.

Before writing a single line of code, the agent creates:

  1. Scope — What are we building?
  2. Tasks — What steps will get us there?
  3. Requirements — What constraints must we respect?
  4. Test protocol — How will we know it's right?

This is your plan. It evolves as you work.


The Agentic Development Loop

┌──────────────┐
│ Create plan  │  ← Agent Dialog: scope, tasks, requirements, tests
└──────┬───────┘
       │
       ▼
┌──────────────┐
│ Execute step │  ← Build one thing
└──────┬───────┘
       │
       ▼
┌──────────────┐
│   Evaluate   │  ← Does it match? What did we learn?
└──────┬───────┘
       │
       ▼
┌──────────────┐
│ Update plan  │  ← Reprioritize, add, remove, shuffle
└──────┬───────┘
       │
       └──────────► Next step

After each step, you re-evaluate the plan. Priorities shift. New tasks appear. Others become irrelevant. The plan is alive.


The Agent Tests Its Own Work

The evaluate step isn't just you looking at the screen. The agent actively verifies its own output.

Using Puppeteer MCP, the agent opens a browser, navigates to the prototype, and checks the acceptance criteria — while you watch.

┌──────────────┐
│ Execute step │  ← Agent builds the feature
└──────┬───────┘
       │
       ▼
┌──────────────┐
│ Agent opens  │  ← Puppeteer navigates to the page
│   browser    │
└──────┬───────┘
       │
       ▼
┌──────────────┐
│ Agent checks │  ← Reads text, clicks buttons, checks states,
│  criteria    │     verifies styling, takes screenshots
└──────┬───────┘
       │
       ▼
┌──────────────┐
│ Agent reports│  ← "5 of 6 criteria pass. Error color mismatch."
└──────┬───────┘
       │
       ▼
┌──────────────┐
│ You decide   │  ← Fix it? Update spec? Move on?
└──────────────┘

The agent narrates what it finds:

"Opening signup form at localhost:3000/signup..." "Headline says 'Start in 2 minutes' — matches spec. ✓" "Clicking submit with empty fields..." "Error message: 'Please enter a valid email' — matches spec. ✓" "Error color is #EF4444 — spec says #DC2626. ✗ Mismatch." "Submit button touch target: 48px — meets 44px minimum. ✓"

The agent handles the measurable criteria. Does the text match? Is the touch target big enough? Does the error state trigger?

You handle the qualitative judgment. Does the flow feel natural? Is the visual hierarchy right? Would a real user understand this?

Together, nothing gets missed. Functional testing happens during development — not after.


Everything Is Logged

Every step is logged. Every action is noted. Every decision is recorded.

This means you can restart the conversation at any time. New session? Load the dialog. The agent picks up where you left off.

docs/F-Agent-Dialogs/
└── 2026-02-10-signup-form/
    ├── dialog.md           ← Plan, steps, status
    ├── decisions.md        ← What was decided and why
    └── changes.md          ← What changed from spec

Who Does What?

Both Freya and Idunn can do agentic development:

Agent Activities When to use
Freya [P] Prototyping, [A] Analysis Design exploration — dream up visuals, components, page layouts. Investigate and understand existing products.
Idunn [D] Development, [F] Bugfixing, [E] Evolution, [R] Reverse Engineering, [T] Acceptance Testing Implementation — generate working code from specifications. Fix bugs, add features, test against specs.

The process is the same. The domain differs.


What You'll Learn

Lesson 1: The Development Agent Dialog

How the process is organized. Creating plans, logging steps, structuring sessions so you can restart at any time.

Lesson 2: Evaluation and Feedback

After each step: evaluate, reprioritize, shuffle. How to give effective feedback to AI agents and keep the work on track.

Lesson 3: When You Get Stuck

Sometimes the agent makes change after change and can't find the problem. When to troubleshoot, when to pivot, and when to ask a developer for help.

Lesson 4: Choosing Your Code Format

Output formats, technology stacks, component libraries, and how to choose between static prototypes and executable code.

Lesson 5: Git Discipline

Keeping your work safe with version control. When to branch, how to commit, working with your team, and letting the agent handle Git for you.


Common Mistakes

Mistake Fix
Jumping in without a plan Start with Agent Dialog every time
Not logging steps Everything gets logged, no exceptions
Keeping a stale plan Re-evaluate after every step
Fighting the AI for too long Know when to ask for help
Building everything at once One step at a time

Practice

Pick one specification and run the full loop:

  1. Create an Agent Dialog with scope, tasks, and test protocol
  2. Execute one step
  3. Evaluate the result
  4. Update the plan
  5. Repeat

Lessons

Lesson 1: The Development Agent Dialog

How the process is organized

Lesson 2: Evaluation and Feedback

Staying in control as the agent builds

Lesson 3: When You Get Stuck

Because AI doesn't always get it right

Lesson 4: Choosing Your Code Format

Understanding what the agent produces and which technology fits your needs

Lesson 5: Git Discipline

Keeping your work safe, organized, and shareable


Tutorial

Tutorial 14: Your First Agentic Session

Hands-on guide to running a complete agentic development loop


Next Module

Module 15: Visual Design →

Add soul and visual polish to your prototypes.


Part of the WDS Course: From Designer to Linchpin