Vibe Coding and Agentic Engineering Are Converging

Simon Willison Sounds the Alarm on AI Coding Convergence

Simon Willison, the prolific open-source developer and creator of Datasette, has raised an uncomfortable observation: vibe coding and agentic engineering — two paradigms most developers consider fundamentally different — are starting to converge in practice. The revelation came during his appearance on Heavybit's High Leverage podcast, Episode 9, hosted by Joseph Ruscio, where a live conversation pushed Willison to articulate something he had been sensing but hadn't yet put into words.

The implications are significant for every developer currently building with AI coding tools. What was once a clear distinction between casual, trust-the-AI experimentation and rigorous, production-grade AI-assisted development is becoming dangerously blurry.

Key Takeaways

• Vibe coding — the practice of letting AI generate code without deeply reviewing every line — is creeping into serious engineering workflows
• Agentic engineering, where AI agents autonomously plan and execute multi-step coding tasks, shares more DNA with vibe coding than practitioners want to admit
• The convergence is happening organically, even among experienced developers who believe they maintain strict review practices
• Willison's self-awareness about this shift highlights a broader industry blind spot
• Current AI coding tools like Claude Code, Cursor, GitHub Copilot, and Devin are accelerating this merger
• The distinction between 'responsible AI coding' and 'vibes-based coding' may be more of a spectrum than a binary

What Vibe Coding Actually Means in 2025

Vibe coding is a term popularized by Andrej Karpathy in early 2024. It describes a workflow where developers describe what they want in natural language, let an AI generate the code, and then run it without thoroughly reviewing every line. The developer 'vibes' with the output — if it works, ship it.

Originally, the concept carried a somewhat playful, even dismissive connotation. It was fine for prototypes, weekend projects, and throwaway scripts. Serious engineers, the thinking went, would never rely on vibes for production code.

But Willison's podcast conversation reveals a more nuanced reality. Even disciplined developers find themselves gradually relaxing their review standards as AI-generated code consistently works. The feedback loop is seductive: the more the AI gets right, the less you feel compelled to scrutinize every function.

This behavioral drift is not a character flaw — it's a natural human response to increasingly capable tools. When Claude or GPT-4o produces correct, well-structured code 90% of the time, the cognitive cost of reviewing every line starts to feel disproportionate.

Agentic Engineering Was Supposed to Be Different

Agentic engineering emerged as the 'serious' counterpart to vibe coding. In this paradigm, AI agents don't just autocomplete or generate snippets — they autonomously plan multi-step tasks, write tests, debug failures, and iterate on solutions. Tools like Claude Code, Devin, and OpenAI's Codex CLI represent this approach.

The key promise of agentic engineering is structured autonomy. The AI operates within guardrails: it writes tests before implementation, follows architectural patterns, and produces code that can be reviewed in meaningful chunks. Unlike vibe coding, agentic engineering supposedly maintains engineering rigor.

But here's Willison's uncomfortable insight: in practice, when an AI agent executes a 15-step plan, writes 500 lines of code, generates tests, and all the tests pass — how carefully does the developer actually review the output? The answer, increasingly, is 'not as carefully as they'd like to admit.'

The structural guardrails of agentic engineering create a false sense of security. Passing tests feel like validation, even though the AI wrote both the code and the tests. This is vibe coding wearing a lab coat.

The Spectrum Problem Nobody Wants to Discuss

Willison's realization points to a fundamental spectrum problem in AI-assisted development. The industry has drawn a sharp line between responsible and irresponsible AI coding practices, but reality refuses to cooperate with clean categories.

Consider the typical progression of a developer using AI coding tools:

• Stage 1: Carefully review every line the AI generates, treating it like a junior developer's pull request
• Stage 2: Skim the generated code, focusing on architecture and logic flow rather than syntax
• Stage 3: Focus primarily on whether the output works (tests pass, feature functions correctly)
• Stage 4: Describe the desired outcome, run the agent, check the results, and move on
• Stage 5: Full vibe coding — trust the AI, ship the output

Most developers who consider themselves agentic engineers are operating somewhere between stages 2 and 4. Very few maintain the discipline of stage 1 consistently, especially under deadline pressure.

This progression is not hypothetical. It mirrors the adoption curve of every productivity tool in software engineering history, from IDEs to Stack Overflow to autocomplete. Humans naturally optimize for speed once trust is established.

Why This Convergence Matters for the Industry

The merging of vibe coding and agentic engineering has concrete implications for how companies build software, how teams evaluate code quality, and how the $15+ billion AI coding tools market evolves.

For engineering leaders, the convergence means that code review processes designed for human-written code are inadequate. When an AI agent generates a complete feature — including tests — traditional PR review becomes theater. New review paradigms are needed, potentially involving adversarial AI reviewers or formal verification tools.

For tool builders, the convergence suggests that the market distinction between 'professional' and 'casual' AI coding tools may collapse. Products like Cursor and Windsurf already span both use cases. The winners will be tools that make the vibe-to-rigor transition seamless, not tools that pretend the boundary is firm.

For individual developers, Willison's honesty is a wake-up call. Self-awareness about where you fall on the vibe-to-agentic spectrum is the first step toward making deliberate choices about when to trust and when to verify.

Key risks of the convergence include:

• Security vulnerabilities introduced by AI-generated code that passes functional tests but contains subtle flaws
• Technical debt accumulating faster than teams realize, as AI-generated code lacks the contextual understanding a human brings
• Skill atrophy among developers who stop reading code critically
• Liability questions when AI-generated production code causes failures
• False confidence from AI-written test suites that achieve high coverage but miss edge cases

How Experienced Developers Are Responding

Willison's approach — publicly acknowledging the convergence rather than pretending it isn't happening — represents one emerging strategy. Transparency about AI reliance allows teams to build appropriate safeguards.

Some practitioners are adopting what might be called 'structured skepticism': using AI agents for generation but maintaining human-driven architectural decisions and security reviews. This approach accepts that line-by-line review of AI output is unsustainable while preserving human judgment at critical decision points.

Others are investing heavily in automated verification layers — separate AI systems that review and challenge the primary agent's output. Companies like Snyk and Semgrep are already positioning their security scanning tools as essential companions to AI coding assistants.

The most pragmatic response may be accepting the convergence and designing workflows around it. Rather than fighting the natural drift toward vibe coding, teams can build systems that make the 'vibe' safer: comprehensive CI/CD pipelines, property-based testing, runtime monitoring, and canary deployments that catch issues before they reach all users.

Looking Ahead: The New Reality of AI-Assisted Development

Willison's podcast revelation is likely just the beginning of a broader industry reckoning. As AI coding tools grow more capable — Claude 4 and GPT-5 are expected later in 2025, promising even more autonomous coding abilities — the pressure to trust and ship AI output will only intensify.

The developers and organizations that thrive will be those who acknowledge the convergence honestly and build systems accordingly. Pretending that agentic engineering is categorically different from vibe coding is increasingly untenable. The real question isn't whether you're vibe coding — it's whether your safety nets are strong enough for the vibes you're riding.

Willison's willingness to surface this uncomfortable truth publicly, through conversations like the High Leverage podcast, is exactly the kind of honest discourse the AI development community needs. The tools are evolving faster than our practices, and closing that gap requires admitting where we actually stand — not where we'd like to be.