AI-Native ডেভেলপারের ৭টা লেভেল

You consult AI in a chat and copy code back by hand.

A strong classic engineer who has tasted AI — but only as a chat buddy. No agents, no repo integration. The project lives in your head, not in the agent's context. This is the starting line of the reform, not a bad place to be.

• ▹You paste code in and out of ChatGPT / Claude web
• ▹Up to ~50% of code touches AI, but as copy-paste
• ▹No MCP, skills, or plan-before-code
• ▹You work solo on your task — deep but narrow

AI writes the code; you still check every line by hand.

You've crossed into AI-native. Code basically isn't written without AI now — but in assistant mode, with manual review of every change. You're learning to phrase the task and delegate the routine.

• ▹~100% of code goes through AI, assistant-style
• ▹You work mostly in one chat / session
• ▹You verify every change manually
• ▹You've wired your first MCP

The agent is your main production mechanism — with a plan and verification.

Agents write the routine, not you. You plan before you code, keep project memory in the repo, and don't take the agent's word for it — you build verification. This is real AI-native middle.

• ▹≥50% of routine code is agent-driven, not pasted
• ▹2+ working MCP, plus skills and plugins
• ▹You keep CLAUDE.md / AGENTS.md current
• ▹Plan-before-code for non-trivial work; you dictate long prompts

You build the AI system for the project — not just the code.

You direct agents at a high level instead of typing routine. A spec becomes production in days. You build reusable harnesses, run parallel agents in worktrees, add evals to CI, and set the safety policy.

• ▹You orchestrate 5–10 parallel agents in worktrees
• ▹Spec → feature in production in days, not weeks
• ▹Reusable skills / harness + evals in CI
• ▹Independent AI review on every meaningful PR; 5+ h autonomous runs

You design the company-wide agent stack and own the AI-native SDLC.

Not just a developer — an architect of the AI delivery platform: model routing policy, cost dashboards, an eval platform, MCP governance and security boundaries for the whole company.

• ▹Company-wide agent stack + model routing policy
• ▹Cost / telemetry dashboards and an eval platform
• ▹MCP governance + an internal skill marketplace
• ▹You run multi-ticket autonomous cycles end to end

You build portable AI methods others adopt, and level people up.

The highest craft tier: you create transferable harnesses and skill-packs that other teams use, embed them into other projects, and raise other developers up the ladder. You define what AI-native means here.

• ▹You create project-agnostic AI methodologies
• ▹Your harnesses / skill-packs are used by other teams
• ▹You embed processes into other projects and support adoption
• ▹You set the company's definition of AI-native

A director, not a coder — one person, full cycle, any artifact.

Roles blur. With agents you take a feature or product through the whole cycle alone — market research → spec → production → promotion — and create artifacts of any kind: code, design, decks, PoC, marketing. The peak of the model.

• ▹Full cycle solo: research → spec → ship → promote
• ▹Artifacts beyond your role: design, decks, marketing, PoC
• ▹≥×3 productivity by covering adjacent functions
• ▹You operate at 'set the task & accept it', not manual execution