Three Approaches: Deep Dive

The Three Approaches to CI/CD & Development

To successfully adopt AI, organizations and developers generally evolve through three working modes. Understanding these distinctions is critical for setting realistic expectations and training goals.

The Evolution of Working Modes

As maturity increases, the responsibility shifts from the Human to the AI.

graph LR
    subgraph "Helper Mode"
    H1[Human drives, AI suggests]
    end
    
    subgraph "Co-creator Mode"
    H2[Human guides, AI builds]
    end
    
    subgraph "AI-as-Software Mode"
    H3[Human defines goals, AI executes]
    end
    
    H1 --> H2 --> H3
    
    style H1 fill:#e3f2fd,stroke:#1565c0
    style H2 fill:#e8f5e9,stroke:#2e7d32
    style H3 fill:#fff3e0,stroke:#ef6c00

1. AI Assisted (Natural Prompting / Zero-Shot)

Status: The industry standard. Most developers are here today.

This is the entry point for almost everyone. It involves using AI tools with natural language prompting (often zero-shot, meaning without providing examples) to get immediate help.

How developers work

Ad-hoc Queries: “Explain this regex,” “Fix this bug,” “Write a function to sort this list.”
Chat-based: Heavy reliance on chat interfaces (ChatGPT, Copilot Chat).
Reactive: AI responds to a developer’s immediate block.

Strengths

Low Barrier to Entry: No new infrastructure or complex prompting skills needed.
Immediate Value: Solves the “blank canvas” problem instantly.
Learning Aid: Excellent for junior developers to understand legacy code.

Limitations

Non-Repeatable: A prompt worked once, but might not work again.
Context Limited: AI often lacks full repository awareness (though tools are improving).
Hallucinations: High risk if the developer doesn’t understand the output.

Skills at this Stage

Skill	Level	Description
Prompting	Basic	Natural language questions, iterating until it works.
Architecture	Traditional	Monoliths or standard microservices.
Testing	Manual	Unit tests written manually, potentially with AI help.
Governance	None	Individual responsibility.

Tools: GitHub Copilot, ChatGPT, Cursor (basic usage), Codeium.

2. AI Driven (Spec-Driven / Vibe Coding + Zero Trust)

Status: Advanced teams and early adopters.

Here, teams start building workflows around AI. Code is no longer the primary artifact—specifications and prompts are.

Vibe Coding & Spec-Driven

Vibe Coding: Rapid, iterative building where the developer focuses on the “feel” and functionality, letting AI handle the implementation details.
Spec-Driven: Writing detailed markdown specifications/prompts that an AI agent uses to generate full files or modules.

Zero Trust AI

Because the AI is generating significant code volume, humans must adopt a Zero Trust mindset.

Trust, but verify: Every AI output is treated as “untrusted” until it passes automated tests and review.

flowchart LR
    Spec[Write Spec] --> Gen[AI Generate]
    Gen --> Review[Human Review]
    Review --> Test[Auto-Test]
    Test -->|Fail| Gen
    Test -->|Pass| Deploy

New Practices

Prompt Libraries: Sharing effective prompts across the team.
Spec Templates: Standardizing how features are requested.
AI Code Review: Using AI agents to review human code before a human sees it.

Skills at this Stage

Skill	Level	Description
Prompt Engineering	Intermediate	Few-shot prompting, chain-of-thought, structuring context.
Spec Writing	High	Ability to describe technical requirements clearly in English.
Code Review	Critical	Shifting focus from “writing” to “auditing” code.
AI Governance	Emerging	Guidelines on what data can be sent to LLMs.

Tools: Copilot Workspace, LangChain, Semantic Kernel, Playwright (AI assisted), Cursor (Composer mode).

3. AI Native (AI as the Product)

Status: The destination. Top tier tech companies and startups.

This stage is fundamentally different. It requires strong software engineering combined with AI engineering. The goal is to build applications where AI is the core engine, not just a dev tool.

AI/LLM Engineering

Agentic Systems: Building systems where multiple agents collaborate (e.g., Planner, Executor, Reviewer).
RAG Pipelines: Architecting complex retrieval systems for domain knowledge.
Orchestration: Managing the flow of data and decisions between models and deterministic code.

graph TD
    User --> Interface
    Interface --> Orchestrator
    Orchestrator --> Agent1[Search Agent]
    Orchestrator --> Agent2[Reasoning Agent]
    Orchestrator --> Agent3[Action Agent]
    Agent1 <--> VectorDB
    Agent3 --> API[External APIs]

Skills at this Stage

Skill	Level	Description
Software Engineering	Expert	Deep understanding of distributed systems, latency, and caching.
LLMOps	Advanced	Evaluations (Evals), monitoring, fine-tuning, versioning prompts.
AI Architecture	Advanced	Designing for non-deterministic behavior and failovers.

Tools: LangGraph, Azure OpenAI, Vector Databases (Pinecone, Weaviate), Langfuse, Kubernetes.

Transformation Summary

Feature	AI Assisted	AI Driven	AI Native
Primary Input	Code Snippets	Specs / Prompts	Strategic Goals
Primary Action	Writing / Debugging	Reviewing / Orchestrating	Architecting / Tuning
Output Volume	Lines of code	Modules / Features	Autonomous Agents
Key Metric	WPM (Words per minute)	Velocity / Throughput	Success Rate / Accuracy