This Is NOT Vibe Coding
Understanding what we actually do.
The Term "Vibe Coding"
Andrej Karpathy coined this term in February 2025. His definition was specific:
"There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."
Key characteristics:
- "Forget that the code even exists"
- For throwaway weekend projects
- Accept AI output without deep understanding
- "YOLO and see if it works"
This is NOT what we do.
Vibe Coding vs Production AI-Assisted Development
| Aspect | Vibe Coding (Karpathy) | AgentOps (Production) |
|---|---|---|
| Philosophy | "Forget code exists" | Understand deeply, delegate implementation |
| Use case | Throwaway projects | Production systems |
| Validation | Accept AI output | Validate ruthlessly |
| Understanding | Optional | Mandatory |
| Debugging | YOLO | Structured recovery |
| Timeframe | Weekend hacks | Long-term maintenance |
| Quality bar | "Works enough" | Production-grade |
| Target | Individual experimentation | Enterprise operations |
The Terminology Problem
When Gene Kim and Steve Yegge published Vibe Coding: Building Production-Grade Software With GenAI, Simon Willison pointed out the disconnect:
"The book carries the subtitle 'Building Production-Grade Software With GenAI, Chat, Agents, and Beyond'—exactly what vibe coding is not."
He's right. The term has been stretched beyond its original meaning.
We use different terminology:
- Production AI-Assisted Development — AI helps, humans validate
- Spec-Driven AI Development — Plans first, execution second
- AgentOps — Operational discipline for AI agent workflows
What We Actually Practice
The Head Chef Model
From Gene Kim's framework, we practice the "Head Chef" approach:
| Role | Description |
|---|---|
| Line Cook | Writes code by hand (traditional development) |
| Vibe Coder | Accepts whatever AI produces (dangerous) |
| Head Chef | Orchestrates AI as "sous chefs," maintains quality standards |
A Head Chef:
- Defines what needs to be built (specs, tests)
- Delegates implementation to AI agents
- Validates every output meets standards
- Understands every line of code shipped
- Maintains production responsibility
"Delegation of implementation doesn't mean delegation of responsibility."
The Three Developer Loops
We operate across three timescales, each with distinct practices:
Inner Loop (Seconds to Minutes)
Direct AI collaboration in immediate coding flow.
| Practice | Why |
|---|---|
| Checkpoint frequently | Recovery options |
| Run tests yourself | Don't trust AI claims |
| Verify AI work | "Tests Passing Lie" pattern |
| Git mastery | Rollback capability |
Middle Loop (Hours to Days)
Multi-agent coordination and extended sessions.
| Practice | Why |
|---|---|
| Written specifications | AI can't read minds |
| Context management | 40% rule (context degradation) |
| Intentional coordination | Prevent workspace collision |
| CI/CD gates | Automated quality enforcement |
Outer Loop (Weeks to Months)
Long-term architecture and organizational patterns.
| Practice | Why |
|---|---|
| API stability | Don't "torch bridges" |
| Modularization | Manage complexity |
| Dev/prod parity | Production safety |
| Fast lane for low-risk | Organizational agility |
Our Quality Standards
Code Understanding
Every piece of code we ship, we can:
- Explain what it does
- Modify without AI assistance
- Debug when it fails
- Maintain long-term
If we can't understand it, we don't ship it.
Validation Gates
Every change passes through:
- AI generates — First draft from AI
- Human reviews — Understanding verified
- Automated tests — Correctness validated
- Security scan — Vulnerabilities checked
- Manual verification — Production readiness confirmed
Failure Recovery
When things go wrong (and they will), we have:
- Documented rollback procedures
- Tested recovery paths
- Blameless postmortems
- Institutional memory updates
Why The Distinction Matters
The critics are right about vibe coding (Karpathy's definition):
- It doesn't scale
- It produces unmaintainable code
- It degrades skills
- It introduces security issues
They're critiquing the wrong target when they apply these criticisms to:
- Production AI-Assisted Development
- Spec-Driven AI Development
- AgentOps methodology
Different practices, different outcomes.
The Core Difference
Vibe Coding: AI writes code, human accepts it
AgentOps: Human designs, AI implements, human validates, system enforces
It's the difference between:
- Copying from Stack Overflow without understanding
- Using Stack Overflow as a starting point, then understanding and adapting
Both use external assistance. One is professional. One isn't.
Next
Our Response → — Why we built AgentOps despite the valid criticism
Navigation
- What The Critics Say — The research and skepticism
- Our Response — Why AgentOps exists
- Back to Home