Senior Devs Got 19% Slower But Believed They Were Faster — The Truth About Vibe Coding

#ai #vibecoding #productivity #agenticengineering

In February 2025, OpenAI co-founder Andrej Karpathy fired off a tweet.

"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."

He was using voice interfaces to tell AI "decrease the sidebar padding by half," pressing "Accept All" without reading diffs, and copy-pasting error messages back to the AI with no commentary. He acknowledged "the code grows beyond my usual comprehension." It was meant for throwaway weekend projects.

The tweet hit 4.5 million views. Searches spiked 6,700% that spring. Collins Dictionary named it Word of the Year 2025. Karpathy later called it "a shower-thoughts throwaway tweet."

One year later, February 2026, Karpathy spoke again. "Vibe coding is now passé." His new term: "Agentic Engineering."

Two Definitions at War

The term "vibe coding" carries two conflicting meanings.

The first is Karpathy's original framing. Accept AI-generated code without understanding it. "Forget that the code even exists." Suitable for prototypes and throwaway projects. Risky for production.

The second is the practical reality. AI generates code, but humans review, test, and understand it. The AI acts as a "typing assistant" or "pair programmer." Production-viable.

Programmer Simon Willison drew the definitive line: "If an LLM wrote every line of your code, but you've reviewed, tested, and understood it all, that's not vibe coding in my book — that's using an LLM as a typing assistant."

The Numbers: Already Mainstream

By 2026, AI-assisted coding is the norm, not the exception. 92% of US developers use AI coding tools daily. 82% of global developers use them at least weekly. 41% of all code written globally is AI-generated. 87% of Fortune 500 companies have adopted at least one vibe coding platform.

The experience-level split is dramatic. Senior developers (10+ years) report 81% productivity gains. Mid-level developers (3–10 years) see 51% faster task completion but need more review time. Junior developers (0–3 years) show mixed results, with 40% admitting they deploy code without full understanding.

The Uncomfortable Research

METR, an AI evaluation organization, ran a randomized controlled trial in July 2025. Experienced open-source developers using AI coding tools were 19% slower. They had predicted they'd be 24% faster. After the experiment, they still believed they'd been 20% faster.

They actually got slower but believed they got faster. The reason: on non-trivial tasks, the cognitive overhead of reviewing, verifying, and fixing AI-generated code negated or reversed the speed benefits.

Security issues are equally real. In May 2025, 170 out of 1,645 web apps generated by the Swedish vibe coding platform Lovable had security vulnerabilities exposing personal information. A December 2025 CodeRabbit analysis of 470 open-source pull requests found AI-co-authored code contained approximately 1.7x more "major" issues compared to human-written code.

Linus Torvalds Got It Right

In January 2026, Linux creator Linus Torvalds used Google Antigravity to vibe-code a Python visualizer for his AudioNoise hobby project. He wrote it in the README: "The Python visualizer tool has been basically written by vibe-coding."

The crucial detail: he hand-coded the C components himself. Vibe coding was for peripheral tooling, not core logic. At an Open Source Summit, he said vibe coding was fine for getting started but a "horrible idea" for maintenance.

Core C code (kernel logic) — hand-coded. Peripheral Python tool (visualization) — vibe-coded. That's the most practical approach in 2026.

2026: From Vibes to Agentic Engineering

Karpathy's February 2026 declaration sets the direction clearly. "Today, programming via LLM agents is increasingly becoming a default workflow for professionals, except with more oversight and scrutiny. The goal is to claim the leverage from the use of agents but without any compromise on the quality of the software."

The evolution: from vibe coding (accepting code without understanding) to agentic engineering (agents generate, humans supervise and verify). The agent patterns from earlier — planning, tool use, reflection — now apply to coding workflows. Agents plan, implement, and test. Humans review, direct corrections, agents revise.

The best practitioners are senior engineers who understand architecture, catch bad AI output, and know when to intervene. They use AI to skip tedious parts while maintaining control over decisions that actually matter.