BMAD-METHOD/docs/learn/module-14-agentic-development/lesson-01-iterative-buildin...

4.9 KiB

Module 14: Agentic Development

Lesson 1: The Development Agent Dialog

How the process is organized


Every Session Starts with a Plan

When you ask Freya or Idunn to develop something, the first thing the agent does is create an Agent Dialog.

This isn't documentation after the fact. It's the plan before anything is built.

The agent reads your input — a specification, a concept, an inspiration image, a rough idea — and creates:

  1. Scope — What are we building? What's the boundary?
  2. Tasks — Ordered steps to get there
  3. Requirements — Constraints we must respect (brand, accessibility, technical)
  4. Test protocol — How we'll verify each step is done right

Then the agent starts executing. One step at a time.


The Agent Dialog Document

# Agent Dialog: Signup Form

## Meta
- Date: 2026-02-10
- Input: P02-signup-form.md (specification)
- Agent: Idunn
- Status: In Progress

## Scope
Build the signup form as a working prototype from the specification.
Includes: form layout, validation, error states, loading states.
Excludes: backend integration, email verification flow.

## Tasks
1. [x] Create form layout with all fields
2. [x] Add real-time validation
3. [ ] Implement error states
4. [ ] Add loading and success states
5. [ ] Responsive adjustments
6. [ ] Accessibility pass

## Requirements
- Must match specification P02 exactly
- Inter font, #0066FF primary
- Mobile-first, max-width 400px
- WCAG AA contrast compliance

## Test Protocol
- [ ] All text matches spec
- [ ] All states render correctly
- [ ] Tab navigation works
- [ ] Touch targets 44px+
- [ ] Contrast ratios pass

## Log
### Step 1: Form layout (completed)
- Created form with email, password, submit
- Matches spec layout
- Decision: Used flexbox for vertical stacking

### Step 2: Validation (completed)
- Added blur-based validation
- Decision: Blur instead of keypress (less jarring)
- Spec impact: Updated storyboard to clarify trigger

### Step 3: Error states (in progress)
- ...

Why Start with the Dialog?

It forces clarity

Before building, you know exactly what you're building and how you'll know it's done.

It enables restart

AI conversations are ephemeral. The dialog is permanent. If you lose context, start a new session, or come back next week — load the dialog. Everything is there.

It tracks decisions

"Why did we use blur validation?" The answer is in the log, not in someone's memory.

It shows progress

At any point you can see: what's done, what's next, what changed.


The Log Is the Heartbeat

As you execute each step, the agent logs:

  • What was built — the output
  • Decisions made — and the reasoning
  • Spec impacts — did anything in the spec change?
  • Issues found — unexpected problems

This log is what makes the dialog a living document, not a static plan.


Plans Change

After completing a step, you and the agent re-evaluate:

  • Did this step reveal something new?
  • Should priorities change?
  • Do we need to add tasks?
  • Should we remove tasks that are no longer relevant?
  • Did the scope shift?

Example: You planned to build the password strength indicator. But after implementing the form, you realize the field is too small to show it on mobile. New task: rethink the indicator placement. Old task: build indicator as-is → removed.

The plan is not a contract. It's a compass.


Types of Agentic Sessions

The same Agent Dialog structure works for all types:

Type Example Agent
Concept exploration "Dream up a dashboard layout for dog trainers" Freya
Proof of concept "Can we do infinite scroll with this data model?" Idunn
Prototype "Build a working signup form from this spec" Idunn
Design inspiration "Generate 3 visual directions for the landing page" Freya
Production code "Implement the complete registration scenario" Idunn

The scope and depth change. The process doesn't.


Session Continuity

Starting a new session

Load the Agent Dialog into your new conversation:

"I'm continuing the signup form implementation.
Here's the current Agent Dialog with plan and log:

[Paste dialog.md content]

We completed steps 1-3. Next is step 4: loading states."

The agent picks up from the log, not from memory.

Handing off between agents

Freya dreamed up a component. Now Idunn needs to build it.

The Agent Dialog travels with the work. Idunn reads Freya's dialog and continues where she left off.


What's Next

In the next lesson, you'll learn how to evaluate results effectively and give feedback that keeps the work on track.


Continue to Lesson 2: Evaluation and Feedback →


← Back to Module Overview

Part of Module 14: Agentic Development