Vibe coding and AI engineers

The AI Engineer role became visible in 2023, followed by “vibe coding” as a named practice in early 2025. The groups overlap, but they bring different levels of software experience and need different forms of education and support.

The AI Engineer

Coined by Shawn “swyx” Wang in his June 2023 essay The Rise of the AI Engineer. The argument:

A new role is forming between traditional ML engineers (who train models from scratch) and traditional software engineers (who don’t touch ML at all).
This role uses foundation-model APIs, prompt engineering, retrieval-augmented generation (RAG), evals, and agentic frameworks to build AI-powered applications.

“A wide range of AI tasks that used to take 5 years and a research team to accomplish in 2013, now just require API docs and a spare afternoon in 2023.”

Andrej Karpathy endorsed the framing publicly shortly after publication, which gave the identity legitimacy.

By 2026, “AI Engineer” is an established career identity:

The AI Engineer Summit (founded by swyx) is one of the most-attended specialist developer conferences.
The Latent Space podcast and newsletter (swyx + Alessio Fanelli) covers the people, tools, and working practices around AI engineering.
The AI Engineer Foundation institutionalises the identity.
Job postings titled “AI Engineer” appear at most developer-product companies; pay tends to be competitive with senior software engineering.

For DevRel teams, AI Engineers are a distinct audience with distinct preferences:

They consume content that assumes familiarity with embeddings, vector databases, RAG, agentic patterns, prompt engineering, evals, and the broader ML application stack.
They evaluate products primarily through hands-on experimentation in notebook environments (Jupyter, Colab, Cursor, Claude Code).
They participate in distinct community spaces (Latent Space Discord, OpenAI Discord, LangChain Discord, Hugging Face Hub).
They are inundated with AI-product marketing and have well-developed filters against it.

Products for AI engineers need examples built around model evaluation, retrieval, agents, latency, and cost. The people teaching those examples need enough experience to discuss the trade-offs.

Vibe coding

Coined by Andrej Karpathy on February 2, 2025 in a public post:

“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.”

In the practice Karpathy described, a developer states an intent, runs the generated code, returns errors to the agent, and iterates without examining much of the implementation.

Karpathy explicitly framed this as appropriate for “throwaway weekend projects” rather than production systems. The framing got picked up far beyond its original scope: by late 2025, “vibe coding” was being used to describe everything from quick scripts to substantial production work, with mixed outcomes.

Audience effects for DevRel

Vibe coding redefines what “developer” means at the margin. Two new audience patterns:

Pattern 1: Existing developers vibe-coding adjacent work.

A backend engineer who vibe-codes a quick admin panel. A frontend developer who vibe-codes a CLI tool. A senior engineer who vibe-codes prototypes for design exploration. This pattern is mainstream; most working developers in 2026 vibe-code at the margin.

For DevRel, this means that more developers evaluate a product by asking an AI agent to build something with it in 30 minutes. Confusing documentation and weak API ergonomics now interrupt that evaluation before a person has read the full quickstart.

Pattern 2: New developers who only vibe-code.

People who never learned to write code in the traditional sense, but who use AI agents to build software. By 2026, this is a real and growing population. Designers building tools. Domain experts building internal apps. Founders building first MVPs.

For DevRel, this is a new audience that needs a different style of engagement:

They don’t read technical reference docs the way traditional developers do.
They evaluate products by whether ChatGPT or Claude can successfully integrate them.
They participate in different communities (Replit, Bolt, Lovable, Cursor-adjacent spaces).
Their failure modes are different (they often can’t debug deeply when the AI fails).

Reports of failed vibe-coded projects describe stripped tests, ignored security advice, leaked secrets, and code the author cannot maintain. Developer products serving this audience need safer defaults and recovery material for those failure modes.

Implications for DevRel content

DevRel content patterns that work for vibe-coders:

Quickstarts that an AI agent can read and faithfully execute (see ./documentation-for-agents.md).
“Build this with AI” tutorial format: explicitly assume the reader is using Claude Code, Cursor, or ChatGPT.
Sample apps that exhibit best practices the AI agent will copy.
Plain-language explanations of concepts that agents tend to get wrong.
Recovery patterns: when the AI agent inevitably produces broken code, what should the developer do?

Most teams in 2026 have not yet figured out the right balance. Teams serving vibe-coders are iterating on quickstarts, sample apps, and AI-readable docs while preserving deeper material for traditional developers.

The JetBrains Human-AI Experience (HAX) research

In April 2026, JetBrains published the Human-AI Experience study analysing two years of telemetry from approximately 800 developers using AI coding assistants. Headline findings relevant to DevRel:

Developers using AI assistants write more code overall.
They spend more than one-third of their time double-checking and editing AI suggestions.
Editing frequency increased substantially among AI users, even though developers themselves perceived minimal change.
Time savings averaged 3.6 hours per week for regular users; daily users showed 60% higher pull-request throughput.
Significant trust gaps remained: even regular users reported incomplete trust in AI-generated code.

The implication: AI assistance reshapes developer behaviour in ways that often elude developers’ own perceptions. Actual behaviour and perceived behaviour diverge. DevRel teams using user surveys to measure their product’s AI-mediated experience get distorted signal; supplementing with telemetry is necessary.

This is the empirical grounding for the wider observation that sentiment about AI productivity tends to overstate measured productivity. DevRel teams that lean only on developer surveys overestimate how well their AI-mediated work is going.

The skill-formation question

Anthropic’s research on how AI assistance affects skill development (early 2026) flagged a genuine concern: AI both accelerates productivity on already-known skills and may hinder the acquisition of new ones.

The skill-formation concern is that a developer may complete an unfamiliar task without learning how to explain or debug the result. Teams can test that concern directly by asking people to modify the code after the assistant is removed.

For DevRel teams at education-adjacent products (JetBrains Academy, courses by Wes Bos / Scott Tolinski, Frontend Masters, etc.), this is a genuine strategic question: how do you teach developers to use AI productively and to retain skill formation? In 2026 the answer is still being worked out.

For DevRel teams at infrastructure / tool companies, the implication is more straightforward: assume that some fraction of your customer-developers will struggle to debug deeply because they have offloaded skill formation to AI. Build your product, docs, and community to compensate: clearer error messages, more accessible community help, more “first principles” content that re-teaches what AI tends to elide.

The cultural texture

A few specific cultural shifts visible in the developer population around AI work:

The end of “I write all my code.” A statement that was identity-defining in 2022 is now rare. Most developers in 2026 acknowledge AI involvement in their work; the variation is in degree.
The new prestige of evals and observability. Among AI Engineers, the ability to evaluate AI output is increasingly the prestige skill. Tools like Braintrust, LangSmith, Helicone, Langfuse are central. DevRel teams in this space talk about evals as much as DevRel teams in 2018 talked about CI/CD.
The “show your prompts” pattern. Developers share prompts the way they used to share dotfiles. Public prompt libraries are a thing. DevRel teams produce sample prompts as deliverables.
The “skill collapse” anxiety. Some senior developers worry that the next generation will not develop the depth they did. Clear teaching material, transparent agent behaviour, and debugging-with-AI guides help people inspect what the agent produced and recover when it fails.

Practical takeaways for DevRel teams

Identify whether your product serves AI Engineers, vibe-coders, traditional developers, or some mix. Each requires different content.
Maintain agent-readable surfaces. Publish complete docs, schemas, examples, and error guidance for the agents developers already use.
Don’t abandon traditional-developer content. The audience is still large, and its depth of trust matters.
Sponsor Latent Space and AI Engineer Summit appropriately. If your product is AI-Engineer relevant, this is the canonical channel.
Track vibe-coding-mediated activation as a separate cohort. It behaves differently from traditional developer activation.
Measure skill formation. Check whether people can explain, debug, and maintain the code they produced with AI assistance.

Primary sources

Shawn Wang, The Rise of the AI Engineer, June 2023 (latent.space / swyx.io).
Andrej Karpathy, post on “vibe coding” via X / personal blog, February 2, 2025.
JetBrains, Understanding AI’s Impact on Developer Workflows, April 2026 (HAX study).
Anthropic, How AI assistance impacts the formation of coding skills, 2026.
Latent Space podcast and newsletter, 2023 to 2026.