A year ago, "vibe coding" felt like the most exciting phrase in tech. Andrej Karpathy coined it, Twitter ran with it, and a generation of developers started shipping apps by describing what they wanted in plain English and letting the AI figure out the rest. It was fast, it was fun, and honestly — for small projects — it worked.
Then people tried to ship it to production.
The results, to put it diplomatically, were educational. Apps that looked polished in demos collapsed under real users. Security vulnerabilities slipped through unnoticed. Codebases became tangles of AI-generated logic that nobody — including the developer who "wrote" it — fully understood. The vibe was there. The engineering wasn't.
Now, in mid-2026, Karpathy has quietly distanced himself from the term he made famous and proposed something more accurate: agentic engineering. And the difference between the two is not just semantic. It is the difference between a hobby and a profession.
What Vibe Coding Actually Was
Vibe coding, at its core, was intent-driven development without structure. You describe what you want. The AI builds it. You run it. If it breaks, you describe the fix. It was fast, low-friction, and deeply appealing to non-technical founders and developers who were tired of fighting syntax.
For prototypes, landing pages, and weekend projects, it remains genuinely useful. The problem was never the tools. The problem was the assumption that the workflow was production-grade.
When you vibe-code, you are not thinking about:
What happens when this scales to 10,000 concurrent users
Whether the authentication logic has an edge case that leaks session tokens
How a developer joining the team in six months will understand what this function is doing
What your database looks like after three months of unreviewed schema migrations
These are not niche concerns for large enterprises. They are basic requirements for any software that real people use for real money.
What Agentic Engineering Actually Means
Agentic engineering keeps the speed of AI-assisted development but adds the discipline layer that vibe coding skipped entirely. Instead of one developer prompting one AI model reactively, agentic engineering means designing a system where AI agents plan, write, test, and ship code — but under structured human oversight at every checkpoint that matters.
The shift has three practical dimensions:
1. Spec-first, not prompt-first. In vibe coding, you describe what you want conversationally and iterate. In agentic engineering, you write a specification first — a structured document that defines behavior, edge cases, failure modes, and acceptance criteria — and then you use AI to execute against it. The spec is the source of truth. The AI is the implementer. You are the architect.
This sounds slow. It is actually faster in the long run, because you stop wasting time regenerating the same features after discovering you built the wrong thing.
2. Human review at system boundaries, not line boundaries. One of the biggest mistakes developers make when using AI tools is reviewing code line by line, the same way they would review a junior developer's pull request. That is the wrong unit of review.
AI-generated code that works at the function level can still be architecturally broken at the system level. The right place for human judgment is at boundaries: where services connect, where data changes ownership, where authentication and authorization decisions are made, where external APIs are called. Those are the load-bearing walls. Review those with rigor. Trust the AI with the rooms.
3. Tests before code, always. Agentic engineering borrows from test-driven development and makes it non-negotiable. Write the test first. Let the AI write the code that passes it. This is not a new idea — TDD has existed for decades — but it takes on a new urgency when the code is being generated by a system that has no institutional memory of your codebase, your users, or the failure that cost you three hours at 2 AM six months ago.
Tests are not just quality gates. In agentic workflows, they are the specification made executable. They are how you communicate constraints to an AI that cannot read your mind.
The Real-World Numbers
This is not theoretical. Organizations that have moved to structured agentic workflows are reporting results that make the earlier hype look modest — but only because they built the foundation correctly.
Stripe's internal AI agents are producing over 1,000 merged pull requests per week. TELUS saved more than 500,000 hours of developer time through AI-assisted workflows. Zapier hit 89% AI adoption across their entire engineering organization. These are not companies that told their developers to "just prompt better." They built platform infrastructure, review processes, security gates, and specification workflows before they scaled up AI usage.
The companies that skipped that foundation are the ones whose names appear in security post-mortems.
What This Means If You're a Developer Today
If you have been doing vibe coding and enjoying it, you are not in the wrong. You are one step away from doing it well. The step is structure.
Add a spec phase before you prompt. Define what done looks like before you ask the AI to build it. Write tests that would tell you if the AI built the wrong thing. Review at the system boundary, not the line level. And when you find a bug in AI-generated code, do not just fix it — understand why the AI made that choice, because it will make the same choice again.
Agentic engineering is not harder than vibe coding. It is more intentional. And in a world where AI can generate a working application in minutes, intentionality is the only competitive advantage that cannot be automated.
Comments (0)
Login to post a comment.