AI engineers declare vibe coding officially dead

The honeymoon is over for vibe coding. Swyx, the influential AI engineering thought leader, declared it dead just months after it began, tweeting "RIP vibe coding 2025-2025" as professional engineers revolt against the slop and security nightmares created by non-technical workers throwing half-baked AI prototypes over the wall. Meanwhile, he reveals code AGI will arrive in 20% of the time of full AGI while capturing 80% of its value, and agent labs like Cognition are now worth more than model labs as even OpenAI admits defeat on building products.

"RIP vibe coding 2025-2025" - Swyx declares it dead as engineers revolt against amateur code. Code AGI arrives 5x faster than regular AGI. OpenAI admits defeat on products.

Engineers revolt as vibe coding creates unfixable messes

Professional software engineers are reaching breaking point with vibe coding, the practice of using AI to generate code through natural language that exploded after Andrej Karpathy's February tweet. Swyx explained the crisis: non-technical workers vibe code something in an hour, then dump it on engineers expecting "the full thing by Friday" without understanding they've only painted a superficial picture missing all the hard parts. The infrastructure layers have specialized so completely for non-technical users that when handoff happens, engineers must rebuild everything from scratch because vibe coders use entirely different tech stacks than production systems.

The inter-engineer warfare is even worse. Some engineers vibe code irresponsibly, leaving security holes and unmaintainable messes for colleagues to clean up. When LLMs hit rabbit holes—which they frequently do—engineers who don't understand the generated code can't debug it. They're "washing their hands" of responsibility while dumping broken pull requests on teammates. The backlash is so severe that engineers are actively searching for vibe coding's replacement, with "spec-driven development" emerging as the leading candidate where humans maintain control and understanding rather than blindly trusting AI outputs.

The timing couldn't be worse for the vibe coding ecosystem. Claude Code launched in March and became a $600 million business, Cursor and Cognition reached unicorn status, but now their target market of professional developers is revolting. Swyx notes everyone he talks to is "sick and tired of vibe coding," with the term becoming synonymous with amateur hour and technical debt. The tools that democratized coding are now being blamed for destroying code quality across the industry, forcing a reckoning about whether making everyone a "coder" was actually a good idea.

Code AGI arrives faster than real AGI with 80% of value

Swyx's bombshell thesis claims code AGI will be achieved in 20% of the time needed for full AGI while capturing 80% of its economic value, making it the most important bet in technology. Code is a verifiable domain where the people building models are also the consumers, creating a virtuous cycle that's already visible. The flexibility of code means these agents generalize beyond coding—Claude Code is already being used for non-coding tasks, with Claude for Excel launching this week built entirely on the Claude Code foundation. The agents being built for coding will become the foundation for all other AI agents.

The evidence is overwhelming: every major AI success story this year involves code. Replit struggled for two years building AI products with no traction, then built a coding agent and hit $300 million revenue. Notion's serious move into agents transformed their business. The pattern is so clear that Swyx joined Cognition, which just acquired Windsurf for a rumored $300 million after Google poached its leadership. He believes coding agents will reach human-level capability years before general AI, and the companies building them will capture most of the value from the entire AI revolution.

This isn't just about making programmers more productive—it's about code becoming the universal interface for AI to interact with the world. Every business process, every automation, every intelligent system ultimately reduces to code execution. The companies that perfect coding agents first will own the infrastructure layer for all AI applications. Swyx's bet is that by the time AGI arrives, code AGI companies will have already captured the market, making general intelligence economically irrelevant for most use cases.

Agent labs overtake model labs as OpenAI gives up on products

Swyx declares vibe coding dead as engineers revolt. Code AGI captures 80% of AGI value in 20% time. OpenAI gives up on products as agent labs dominate.

The AI industry is bifurcating into model labs that build foundation models and agent labs that build products, with agent labs suddenly winning. OpenAI's Sam Altman essentially admitted defeat yesterday, saying "we're giving up on products" and will focus on being a platform where third parties "should make more money than us on our models." This shocking reversal proves Swyx's thesis that shipping products first beats shipping models first. While model labs raise money, hire researchers, buy GPUs, and disappear for months, agent labs like Cognition ship working products immediately and iterate based on user feedback.

The swim lanes are now crystal clear: join a model lab to work on AGI, join an agent lab to build products that actually serve users. Model labs treat applied engineers as second-class citizens, paying them half what researchers make. At Meta, being an applied AI engineer is "low status" compared to research roles. Meanwhile, agent labs are reaching astronomical valuations—Cognition at $10 billion, Cursor and others approaching similar heights—by focusing entirely on product-market fit rather than benchmark scores.

The implications for enterprise buyers are massive. They can no longer just deal with OpenAI, Anthropic, and Google, assuming these platforms will build everything. As model labs retreat to infrastructure, enterprises must now evaluate dozens of agent labs building vertical solutions. The procurement process that favored dealing with three vendors is being forced to expand dramatically. Anthropic remains the wild card, with Claude Code functioning as an agent lab within a model lab, but even they're proving that products, not models, capture value in this new era where everyone has access to the same foundation models but only some can build products people actually want.