The Levels and Evolution of AI Coding

The development of AI coding can be compared to the levels of autonomous driving:

• L2 Stage: This stage is already relatively mature. Many people are using tools like GitHub Copilot for code completion, which significantly boosts productivity. However, engineers still need to maintain full control over the process.

  
  

• L3 Stage: We are now moving into the "semi-autonomous" phase. AI can generate larger blocks of code, such as completing functional modules. This stage demands a higher level of collaboration between humans and AI.

  
  

• L4 Stage: This is where AI takes the lead with humans providing assistance. While still in its early research phase, it may take a longer time before becoming a practical reality.

  
  

• L5 Stage: Fully autonomous AI coding. At this stage, platforms would be able to understand complex business requirements, learn domain-specific knowledge, and generate solutions that meet those needs. From requirements to final code, the entire process would be automated without any human input. Such platforms would also have the ability to self-reflect and iterate, supporting multiple programming languages, frameworks, and industry scenarios. This is the ultimate goal, but achieving it will require significant technological advancements.

  
  

Much like the progression from manual to automatic to smart driving, AI coding is undergoing a similar evolution. The goal is not just to make AI better at completing code but to create an entirely new "intelligent control center" to help engineers work more efficiently.

In the timeline of human technological progress: L5 is distant, L4 is premature, L2 is conservative, and L3 is right on time.

A fundamental truth: Large AI models will gradually reach and stabilize at a level comparable to mid-level engineers (with 3-5 years of experience), each having their strengths and weaknesses.

Diverging Views on AI Coding

Currently, there are two polarized views on AI coding:

1. The Conservative Perspective

   
   

   Some believe that AI coding is still in its infancy. Its accuracy is limited, and the generated code often requires significant human adjustments and corrections, making it unlikely to genuinely replace human engineers anytime soon.

   
   

2. The Radical Perspective

   
   

   On the other hand, some, especially non-engineers, are highly optimistic. They think AI coding will completely disrupt the engineering market, leading to a massive wave of unemployment among engineers.

   
   

In reality, AI coding is currently in an "intermediate state." Over the next few years, it will serve as a powerful assistant for senior engineers, significantly accelerating development efficiency rather than entirely replacing them.

One senior engineer shared a thought on X（ex-Twitter) that left a lasting impression:

"Looking back at how I wrote code two years ago, I can’t imagine how I managed to type everything line by line."

Now, code is often written in chunks. In the near future, engineers might look back at today’s coding practices and find them equally archaic.

Thus, if we fail to embrace the changes brought by AI coding in time, we risk becoming so-called “outdated engineers.”

That said, there’s no need to panic just yet. AI coding still has many limitations, and its advancement is gradual, giving us ample time to adapt.

The Optimal Path to L5 AI Autonomous Coding: CDE as the Core Engine of AI-Native Control

Dimension	Plugin-Based Tools	AI-IDE	AI-CDE	Remarks
AI Assistance Level	L2	L2+	L3~L4	Deeper AI integration, e.g., conversational
Environment Control	Weak	Moderate	Strong	Package management, functional testing, etc.
AI Visualization	None	None	Available	Allows humans to observe and modify AI processes
AI Interaction	Highly Limited	Limited	Unlimited	Adjustment capabilities like AI DIFF, plan changes
Implementation Difficulty	🌟	🌟🌟🌟	🌟🌟🌟🌟🌟	High technical requirements
Maturity Timeline	Early	Relatively Early	Later	AI-CDE needs additional editing capabilities

CDE (Cloud Development Environment) is a critical engine in the AI programming ecosystem. Why is CDE so important? Simply put, relying solely on an AI Agent cannot unlock its full potential. Developing an AI Agent is no longer a significant barrier, as many companies can now create their own. However, for these agents to deliver real-world results, they require strong "hands and feet" to execute complex tasks, which is precisely the role CDE plays.

Human	L5 AI Autonomous Programming	Core Capabilities
Brain	LLM (Large Language Model)	Open-source large models + Fine Tuning
Heart	AI Agent For Coding	· LATS RAG · Multi Planning + Reasoning · Short/Long Memory · Search Browse · Smart Context · Console/Shell Auto Tool
Hands & Feet	CDE (Cloud Development Environment)	· Globally Leading: Achieves "Three Zeros" – zero startup, zero switching, and zero latency technical benchmarks. · Full-stack Self-developed: Independent of any native IDE, enabling rapid iteration capabilities. · PaaS Architecture: Rich APIs, born for AI-native development.

CDE: The Shared Workspace for Humans and AI

CDE serves as a shared workspace for humans and AI, combining the functionalities of an operating system, language management tools, package managers, and source code management systems.

If we think of it on a larger scale, it’s like a "full-stack workspace" where humans and AI collaboratively write code. In this system, both humans and AI actively control the development process, with real-time visualization and interaction to complete tasks together.

By contrast, traditional IDEs (Integrated Development Environments) are limited to source code management and editing, unable to handle deeper environmental tasks like OS adaptation, language version management, or dependency configuration. This makes traditional IDEs hit a ceiling when it comes to supporting modern development needs.

Dimension	AI-IDE	AI-CDE
Core Functionality	Intelligent tools for optimizing human efficiency	Deep human–AI collaboration in the development process
Developer Role	Primary driver, AI as an assistive tool	Co-creator, AI as a team member
Scope of AI	Code completion, error detection, local optimizations	Architecture design, module development, full development chain
Interaction Style	Reactive (developer requests help)	Proactive (AI suggests or discusses solutions)
Target Audience	All developers	Teams or complex projects
Autonomy	Low, fully dependent on human guidance	High, capable of completing certain sub-tasks independently

AI Agent and CDE: The Evolution Path by Complexity

The following chart divides development capabilities into five levels (C1 to C5), ranging from single-file handling to supporting massive projects with millions of files.

Currently, Clacky has achieved support for projects with 10,000 files, making it capable of handling real-world projects for small to medium-sized teams. However, reaching C5-level complexity will require further iterations.

Category/Capability	C1 (Single File)	C2 (5 Files)	C3 (100 Files)	C4 (10k Files)	C5 (1M Files)
Example	Single file generation	Small toy projects	Small projects (individual developers)	Real-world projects (small teams)	Large-scale engineering
AI Agent Capability	None	Context building	Summarization	Decomposition	Complex analysis
CDE Capability	None	File tree management	Dependency installation, syntax parsing	Full environment handling	Performance optimization

Clacky Core Architecture Diagram

The diagram below illustrates the overall design of the Clacky system. We place AI and humans on equal footing, working collaboratively to complete development tasks. The backend adopts a headless architecture of the CDE, responsible for supporting core logic, while the frontend represents the interface layer of the CDE, providing users with an intuitive interactive experience. Additionally, the two sides of the diagram showcase Clacky's frontend (built with TypeScript and Next.js) and backend (built with Go Gin) architectures, which together form a complete development process and functional framework.

Clacky’s Position and Features

Clacky is an L3 Agentic AI-CDE designed for serious developers. Its core philosophy is to enable seamless collaboration between humans and AI, offering greater control and flexibility compared to traditional chat-based AI tools. Key features of Clacky include:

• Thread-Based Development: Similar to Git branches, Clacky encourages creating independent threads for each feature. This ensures better organization and aligns with best practices like Pull Requests.

  
  

• Code Generation: From line-by-line suggestions to module-level generation, Clacky accelerates development while keeping developers in control.

  
  

• Full-Stack Environment: Supporting multiple languages and frameworks, Clacky offers pre-configured tools and middleware to streamline onboarding and maximize efficiency.

  
  

• Multi-Agent Collaboration: Clacky provides specialized roles to mimic a professional development team, including:

  
  
  • Analysis Role: Analyzes project goals and technical requirements.
  • Spec Role: Refines requirements into clear objectives.
  • Plan Role: Creates development plans and task prioritization.
  • Code Role: Generates code based on requirements.
  • Check Role: Performs automated code checks to ensure quality.
  • Chat Role: Handles informal discussions and quick problem-solving.
  • Git Role: Manages branches, commits, and PR descriptions.
    

By combining these features, Clacky aims to redefine how developers and AI collaborate in building tomorrow’s software.