Vibe Coding – The Experimental Entry Point to AI Development

Vibe Coding is a term coined by Andrej Karpathy (OpenAI co-founder, former Tesla AI chief) on February 3, 2025. It describes the experimental, ad-hoc approach to AI-assisted software development – developing by feel and intuition, without strict rules or formal processes. The name derives from "coding by vibe": you describe to the AI what you want, and it generates the code.

Vibe Coding has its place as an entry point and for quick prototypes. For professional, productive development, however, the path leads further: through Agentic Coding to Agentic Engineering. Karpathy himself made this transition when he introduced the successor term "Agentic Engineering" in February 2026.

Characteristics of Vibe Coding

1. Ad-hoc Prompts

Prompts arise spontaneously, without prior planning:

• "Give me a login page"
• "This doesn't work, fix it"
• "Add another button"
• No structured description of requirements

2. No Persistent Context

• No CLAUDE.md or .cursorrules file
• Coding conventions are not documented
• Every session starts from zero
• AI behaves differently with every prompt

3. Trial-and-Error Workflow

• Generate code → try it out → if it doesn't work, try again
• No systematic testing
• Bugs are fixed ad-hoc instead of debugged systematically
• No build checks before commits

4. Inconsistent Code Style

• Sometimes TypeScript, sometimes JavaScript
• Different naming conventions
• Different architecture patterns in the same project
• No unified error-handling approach

5. No Documentation

• Decisions are not recorded
• No "why" for architecture choices
• Lessons learned are lost
• The next developer has to figure everything out from scratch

Vibe Coding vs. Agentic Coding

Aspect	Vibe Coding	Agentic Coding
Structure	None	CLAUDE.md, Context Engineering
Consistency	Variable	Uniform
Documentation	Barely present	Memory Files, Session Handover
Testing	Ad-hoc	Systematic after every change
Suitable for	Prototypes, learning, experiments	Production code, team work
Long-term maintainability	Difficult	Good

When Is Vibe Coding Okay?

1. Learning and Experimenting

When getting to know AI assistants:

• "How does React Hook X work?"
• "Show me an example of TypeScript Generics"
• "What does this code do?"
• No stress, no production pressure

2. Quick Prototypes

Throwaway code for proof-of-concepts:

• "I just want to quickly see if the idea works"
• Code is deleted after testing
• No maintenance needed

3. One-Off Scripts

Helper scripts that only run once:

• Data migration that is only performed once
• One-time report or analysis
• Quick-and-dirty automation

4. Personal Projects

Solo projects without a team:

• Hobby project, just for yourself
• No maintainability requirements
• You are the only "user"

When Do You Need Agentic Coding?

⚠️ Switch to Agentic Coding when...

• Team work: Multiple developers work on the code
• Production code: Code runs with real users
• Long-term project: Code must be maintained over months/years
• Critical systems: Bugs have serious consequences
• Scaling: The project grows beyond a few files

The Transition: From Vibe to Agentic

Phase 1: Create CLAUDE.md

1. Create a CLAUDE.md in the project root
2. Document:
   • Tech stack (e.g. "Vue 3 + TypeScript")
   • Most important code conventions
   • Deployment workflow
3. Start with 10–20 rules, no more

Phase 2: Introduce Testing

1. Build check after every change: npm run build
2. Type check: npx tsc --noEmit
3. Linting: npm run lint

Phase 3: Create Memory Files

1. Create MEMORY.md or dokumentation/ folder
2. Document important decisions
3. Record known issues

Phase 4: Establish Session Handover

1. When context is low: write a session summary
2. Document open TODOs
3. Next session can continue seamlessly

Typical Problems with Pure Vibe Coding

1. "It worked yesterday..."

Without tests and versioning it is unclear:

• What was changed?
• Why doesn't it work anymore?
• How do I get back to a working version?

Solution: Git integration, build checks, tests

2. "The AI suggests something different every time"

Without context engineering the AI is inconsistent:

• Sometimes TypeScript, sometimes JavaScript
• Sometimes async/await, sometimes Promises
• Different error handling patterns

Solution: CLAUDE.md with clear rules

3. "I no longer know why we did it this way"

Without documentation, knowledge is lost:

• Architecture decisions not traceable
• Workarounds without explanation
• When the team changes: learn everything again

Solution: Memory files for important decisions

4. "The build is broken, but I don't know since when"

Without build checks, errors accumulate:

• Build fails at deployment
• TypeScript errors pile up
• Debugging becomes time-consuming

Solution: Build check after every change

Vibe Coding: A Learning Phase

Vibe Coding is often the first contact with AI-assisted development. This is completely normal and even recommended as an entry point:

• You learn how the AI "thinks"
• You develop a feel for good prompts
• You discover what is possible
• You make mistakes in a safe environment

But as soon as things get serious – as soon as code goes to production or a team works together – the transition to Agentic Coding is important.

The Evolution: From Vibe Coding to Agentic Engineering

Notably, Karpathy himself moved beyond his own term: exactly one year after "Vibe Coding," he introduced Agentic Engineering in February 2026 – because "Vibe Coding" no longer captured the professional demands of modern AI development. The evolution follows three clear stages:

1. Vibe Coding (Feb 2025): Experimental entry – generating code by feel
2. Agentic Coding (mid 2025): Professional practice – structured context, testing, documentation
3. Agentic Engineering (Feb 2026): Enterprise discipline – multi-agent orchestration, quality gates, architecture focus

Conclusion: Vibe Coding as a Necessary First Step

Vibe Coding is not a "bad" approach – it is the natural entry into a new era of software development. At Elasticbrains, we have completed this transition with over 2,000 person-days of practical experience. In our Agentic Coding Workshop, we guide teams along this path.

Further Resources

• Glossary: Agentic Coding, Agentic Engineering, Multi-Agent System
• Workshop: Agentic Coding Workshop
• Service: AI Development