↓ Skip to main content
sdd.sh
sdd.sh
  1. Articles/

From Vibe Coding to Agentic Engineering: The Paradigm Shift That Outran Its Own Branding

·1493 words·8 mins·
Table of Contents

On February 2, 2025, Andrej Karpathy posted a short X thread that got 4.5 million views. The core idea: “fully give in to the vibes,” let the LLM write all the code, don’t even read what it generates. He called it “vibe coding.”

Collins Dictionary named it Word of the Year 2025. MIT Technology Review named generative coding one of its 10 Breakthrough Technologies of 2026. By most accounts, vibe coding had won the cultural moment.

And then, in early 2026, Karpathy declared it passé and introduced a replacement: agentic engineering.

The year between those two declarations is among the most compressed paradigm shifts in software development history. Understanding why it happened so fast tells you a lot about where AI-assisted coding is actually going.

What Vibe Coding Actually Said
#

The original tweet was short and deliberately provocative. Karpathy wasn’t describing a rigorous methodology — he was naming something developers were already doing informally and mostly pretending not to. The practice: describe what you want, accept what the AI generates, fix errors by describing them to the AI again, never read the code yourself.

“It’s not really coding,” Karpathy wrote. “I just see stuff, say stuff, run stuff, and it mostly works.”

For small prototypes, weekend projects, and throwaway tools, this was liberating. The friction between idea and working software dropped dramatically. Developers who would have spent three days on a data pipeline could have something running in three hours. Non-developers could build tools that previously required hiring engineers.

The marketing machine pulled in two directions simultaneously: AI tool companies amplified the message as validation; developers with 20 years of hard-won expertise quietly winced.

The Backlash Was Also Right
#

By mid-2025, counter-data was accumulating. A December 2025 CodeRabbit analysis of 470 open-source pull requests found that AI co-authored code had 1.7x more major issues, 75% more misconfigurations, and 2.74x more security vulnerabilities than human-authored PRs.

A January 2026 paper — “Vibe Coding Kills Open Source” — made a different argument: that fully surrendering to AI reduced developer engagement with open-source communities, because the act of reading code, contributing to discussions, and understanding implementation details was where community bonds formed. If you never read the code, you never have anything to say about it.

Both critiques landed. Vibe coding, taken literally, produced worse security posture and was arguably hollowing out the engineering culture that made open-source software good in the first place.

But the critiques also missed something. The problem wasn’t AI-assisted coding. The problem was unsupervised AI-assisted coding — humans removing themselves entirely from the loop in contexts where the loop existed for good reason.

Benchmarks and the Reality Check
#

SWE-bench Verified — the benchmark AI coding tools cited endlessly throughout 2025 — has been effectively retired. A Scale AI audit found 59.4% of its hard tasks have flawed tests. OpenAI stopped reporting scores on it after contamination concerns became too significant to ignore.

The replacement, SWE-bench Pro, tells a different story. Its 1,865 long-horizon tasks require an average of 107 lines changed across 4.1 files. Top scores as of early 2026:

  • Claude Opus 4.5: 45.9%
  • GPT-5 (High): 41.8%
  • GPT-5.2 Codex: 41.0%

The same models scoring 80%+ on SWE-bench Verified score 41–46% on SWE-bench Pro. That gap is the gap between “solving isolated problems on known codebases” and “doing real software engineering work in production repositories.” The dominant failure mode for top models on Pro is context overflow — which directly explains why the 1M context window and Compaction API are strategically significant, not just marketing.

Vibe coding at 45% reliability is appropriate for prototypes. It is not appropriate for production infrastructure, security-sensitive code, or systems where failures page an on-call engineer.

The Successor: Agentic Engineering
#

Karpathy’s reframing is worth quoting precisely. His characterization of the shift: “‘agentic’ because the new default is that you are not writing the code directly 99% of the time, you are orchestrating agents who do and acting as oversight.”

The distinction from vibe coding is exact:

  • Vibe coding: surrender to the AI; don’t read the output; trust the vibes
  • Agentic engineering: orchestrate multiple AI agents; act as strategic oversight; review results before shipping

The human role in agentic engineering is closer to engineering management than to typing. You set goals, structure problems for parallel execution, review outputs, and steer based on what you see. You don’t necessarily write every line, but you absolutely read the code — or at least the parts that matter.

This is a better mental model for what the data actually shows. It explains why the security failures in vibe-coded projects happened: there was no oversight layer. It explains why teams that adopted async agent workflows with proper review processes — Claude Code’s Agent Teams, Jules’s VM-isolated execution, OpenAI Codex’s sandboxed cloud agent — saw better outcomes than teams running in pure vibe mode. And it explains why Spec-Driven Development works: the spec is the strategic brief the engineering manager gives the team before work begins.

What the Numbers Actually Show
#

Despite the critiques, adoption is unambiguous. By 2026:

  • 92% of US developers use AI coding tools daily
  • 41% of all code written globally is now AI-generated
  • The AI coding tools market sits at $4.7 billion in 2026, projected at $12.3B by 2027

The argument was never whether AI tools belong in development workflows — they clearly do. The argument is about the posture the developer takes toward them. Vibe coding said “trust the output.” Agentic engineering says “direct the process.”

That shift is commercially important too: the tools built for vibe coding (fast autocomplete, single-agent generation, IDE autocomplete) are different from the tools built for agentic engineering — multi-agent orchestration, async VM execution, long-context memory, MCP-connected external services, structured review workflows.

What This Means for the Tooling Landscape
#

The vibe coding era favored tools with great autocomplete and fast iteration loops. You typed, the AI suggested, you accepted or rejected. Cursor, GitHub Copilot, and Tabnine were well-positioned for that paradigm.

The agentic engineering era favors different primitives:

Async execution: Jules runs in a Google Cloud VM while you do other work; Claude Code Agent Teams spawns up to 15 independent teammates working in parallel; OpenAI Codex executes in sandboxed cloud environments. You assign work and return to results, rather than watching an agent type in real time.

Terminal-native orchestration: Claude Code’s CLI model wins over IDE wrappers because agentic engineering doesn’t require a GUI. Long-running agents, tmux panel layouts for parallel teams, scriptable workflows — these are terminal-native primitives that IDEs weren’t designed to express.

Long-running context: With 1M tokens generally available for Opus 4.6, week-long agent sessions are viable. The context overflow failures that dominated early SWE-bench Pro results are becoming solvable. Compaction means agents no longer hit walls mid-task on large codebases.

Spec-first workflows: Writing goals before generating code is the natural interface for agentic engineering. The developer acts as architect; the agents handle implementation. This is SDD’s core claim, and the 2026 tooling landscape has converged on it: Windsurf’s Plan Mode, Jules’s plan-then-execute model, and Claude Code’s structured task planning all reflect the same underlying logic.

IDE-centric tools like Cursor are adapting — Cursor 2.0’s Plan Mode and parallel agents are direct responses to this shift — but the architectural starting point still anchors them to synchronous, human-in-the-loop workflows. The developer watches agents work. That model doesn’t scale to 15-agent teams or week-long async sessions.

The Branding Outpaced the Practice
#

“Vibe coding” became Word of the Year before “agentic engineering” was a phrase because “vibe coding” described something that felt new and was easy to demo in 30 seconds. “Agentic engineering” describes something that requires setup, thought, and familiarity with multi-agent architectures to appreciate.

But the harder thing is what actually works at scale. Vibe coding was the right mental model for 2025’s tool capabilities. Agentic engineering is the right mental model for 2026’s.

The timeline compressed faster than anyone predicted. In 13 months, the paradigm that Karpathy named became insufficient to describe the paradigm he was helping build. That’s not a critique of the original framing — it’s a measure of how fast the underlying technology moved.

The next question isn’t whether to use AI in development workflows. It’s whether you’re vibe coding (trusting outputs without oversight) or agentic engineering (orchestrating agents with strategic direction). The tools, the benchmarks, and the failure data all point in the same direction.

Sources

Related