Vibe coding and agentic engineering are getting closer than I'd like

The Spark: An "Unsettling" Realization from a Podcast Veteran software engineer Simon Willison candidly shared an "unsettling" discovery in a podcast interview. He had previously drawn a clear line between "vibe coding"—where non-programmers or developers let AI generate code based on feel, without caring about quality—and "agentic engineering," his term for professionals using AI tools to build high-quality systems. However, as AI coding agents have become more reliable, this once-firm boundary has begun to blur in his own work practice. Deconstructing: The "Curse" of Reliability and the Vanishing Review The core shift lies in the increased reliability of AI tools. Willison gives an example: when you ask Claude Code (an AI coding agent) to build a standard JSON API endpoint, it almost always gets it right. This trust in "it just works" has an unintended consequence: he finds himself no longer reviewing every line of AI-generated code destined for production. This creates a strong sense of "guilt." He realizes he is sliding from a professional engineer who "thoroughly considers every line of code" to a "vibe coder" who operates on "if it looks right, it's fine," despite his 25 years of experience. This isn't a regression in skill, but a behavioral shift driven by tool reliability. Trend Insight: The Responsibility Gulf from "Personal Tool" to "Production System" Willison's reflection reveals a deeper trend: AI-assisted programming is reshaping the responsibility model of software engineering. He previously believed "vibe coding" was only acceptable for personal tools (where bugs only hurt yourself), while building software for others necessitated "agentic engineering" with full quality ownership. Now, the power of the tools allows even professionals to unconsciously adopt a "vibe coding" workflow when building production systems. It's like working in a large company where you depend on a service delivered by another team. You don't review that team's every line of code, but you rely on their professionalism and organizational processes. AI agents now play the role of that "other team," but they lack the accountability and communicative context of a human team. Practical Value: Establishing a New "Trust but Verify" Workflow What does this mean for all developers using AI tools? First, a new mental model is needed. AI should no longer be seen merely as a "code completer," but as a "highly reliable but supervised junior developer" or an "external service team." Second, workflows need adjustment: 1. Define Boundaries: For core, complex, security-sensitive modules, code review must be insisted upon. 2. Automate Verification: For standard, repetitive code generated by AI (like CRUD endpoints), trust can be given, but it must be backed by a robust "verification layer" of automated tests, monitoring, and documentation to replace manual review. 3. Embrace the New Normal: In the future, an engineer's core value may shift from "writing every line of code" to "defining problems, designing architecture, setting quality standards, and verifying AI output." Counter-Intuitive Insight The most counter-intuitive point is that tools becoming "better" and "more reliable" might actually decrease the quality of the final output by eroding engineers' ingrained habits of rigorous review. This isn't a technological regression, but a subtle alienation of behavioral patterns in human-AI collaboration. Willison's "unease" is a manifestation of professional responsibility, reminding us that while embracing AI for efficiency, we must be wary of a "quality slide" and consciously build new engineering disciplines to meet this challenge.