Agent Behavior Development (ABD)

Agent Behavior Development (ABD) is a software development paradigm that treats agent behavior as the primary engineering artifact.

In ABD:

• Code is not the first-class citizen.
• Prompts, guardrails, output formats, review rules, and recycle loops are.

This repository does not provide tools. It defines a development paradigm, not an SDK.

---

One-Sentence Definition

Agent Behavior Development (ABD) is a paradigm where humans design and evolve the behavior of agents, and code is a byproduct validated by evidence.

---

graph TB

  subgraph Planning
    A[Product Intent or Sprint Goal]
    B[Notlanot Sprint Plan with Personas]
    C[Task Packets]
    D[Conflict Inputs]
    E[Conflict Scoring and Replan]
  end

  subgraph PromptOps
    F[Prompt Library]
    G[Prompt Compiler]
    H[Guardrails]
    I[Output Format Contract]
  end

  subgraph Execution
    J[Agent Run]
    K[Structured Output]
    L[Evidence Production]
  end

  subgraph ReviewRecycle
    M[Scorecard]
    N[Tags and Anti Patterns]
    O[Prompt Patch Rules]
    P[Promote to Library]
    Q[Recycle Outputs]
    R[Changelog and Governance]
  end

  A --> B
  B --> C
  C --> D
  D --> E

  E --> G
  F --> G
  G --> H
  H --> I
  I --> J

  J --> K
  K --> L
  K --> M
  L --> M

  M --> N
  N --> O
  O --> J
  O --> P
  P --> F

  M --> Q
  Q --> R
  R --> B

  style F fill:#9C27B0,color:#fff
  style M fill:#FF9800,color:#fff
  style O fill:#F44336,color:#fff
  style L fill:#4CAF50,color:#fff
  style I fill:#1565C0,color:#fff

Loading

---

Why ABD Exists

Traditional development paradigms assume:

• Humans write specifications.
• Humans write tests.
• Humans write most of the code.
• Tools assist.

That assumption no longer holds.

Modern agents:

• Propose behaviors
• Generate test strategies
• Produce implementation drafts

Yet teams still:

• Debug code instead of behavior
• Retry prompts instead of fixing agent constraints
• Treat prompt failures as one-off accidents

ABD exists to systematize agent behavior improvement.

---

The Core Shift: Code → Behavior

The defining shift of ABD is simple:

The primary refactoring target is agent behavior, not code.

In ABD, when output quality is poor:

1. Fix the prompt
2. Tighten guardrails
3. Lock output formats
4. Improve review rules
5. Strengthen recycle logic
   Only then adjust code.

---

What ABD Is

ABD is:

• Agent-first, not agent-autonomous
• Behavior-driven, not prompt-driven
• Evidence-oriented, not test-obsessed
• Governance-heavy, not “just try again”

ABD introduces:

• Prompt-as-artifact discipline
• Behavior scorecards
• Mandatory recycle outputs
• Conflict-aware task planning
• Agent behavior retrospectives

---

What ABD Is Not

ABD is NOT:

• AI-assisted TDD
• BDD automation
• “Let the agent do everything”
• Prompt engineering tips
• A replacement for Scrum

ABD complements Scrum by adding an agent behavior governance layer.

---

ABD vs TDD vs BDD

Aspect	TDD	BDD	ABD
Who defines behavior	Human	Human	Agent (under rules)
Primary artifact	Test	Scenario	Prompt + Guardrails
What is refactored	Code	Code	Agent behavior
Failure resolution	Rewrite code	Rewrite scenario	Patch prompt/flow
Evidence	Tests	Scenarios	Tests, checks, demos
Paradigm focus	Code correctness	Business behavior	Behavior production system

TDD and BDD are techniques in ABD, not paradigms.

---

Non-Negotiable Rules (v1)

1. Agent output must follow a fixed format.
2. Evidence must be proposed before implementation.
3. Assumptions must be explicit.
4. Agents may ask up to 3 questions and must stop.
5. Every task must produce recycle outputs.
6. Behavior changes must be versioned.

---

Repository Map

• MANIFESTO.md → Hard rules of the paradigm
• 0_introduction/ → What ABD is and is not
• 1_core/ → Principles and artifacts
• 2_process/ → Lifecycle, sprint and task models
• 3_promptops/ → Prompt operations and metrics
• 4_templates/ → Ready-to-use templates
• 5_examples/ → End-to-end worked examples

---

First-Week Adoption Guide

Day 1:

• Read MANIFESTO
• Use the core prompt template for one task
• Score the output

Day 2–3:

• Run 3 more tasks
• Tag recurring failures

Day 4:

• Patch the prompt
• Document behavior improvement

Day 5:

• Sprint review focused on agent behavior, not features