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

Origin: A Podcast Conversation Sparking Unease

In a recent podcast interview, veteran software engineer Simon Willison shared an observation that even he found "unsettling": the clear line he had long maintained between two modes of AI-assisted programming—"vibe coding" and "agentic engineering"—is blurring in his own practice. This realization matters because it touches on the fundamental shift in the developer's role, responsibility, and trust dynamics as AI becomes deeply embedded in the core software development process.

Breakdown: The Core Distinction and Its Blurring Lines

Willison was an active definer of the "vibe coding" concept. He describes it as a "if it feels right, it is right" approach: a user (who might not even be a programmer) describes what they want to an AI, receives code, and is satisfied if it runs. The process involves no concern for code quality, security, or maintainability. It's like quickly building a personal model with LEGO bricks—if it collapses, you just rebuild it, as long as you're having fun. Its use case is clear: building tools for personal use where issues only affect you.

In contrast, "agentic engineering" is the practice of professional software engineers. Engineers leverage their own expertise (e.g., in security, performance, operations) and use AI tools as capability amplifiers, aiming to build higher-quality production software, not just faster low-quality software. Here, AI is a collaborator, but the engineer remains the architect and reviewer with ultimate responsibility.

The turning point, however, is the significant increase in reliability of AI coding agents (like the Claude Code he mentions). When an AI can consistently and correctly generate standardized code—like "a JSON API endpoint that runs a SQL query"—and automatically add tests and documentation, the seasoned engineer Willison finds he no longer needs to review every line. He begins to treat the AI agent's output like that of a reliable internal team: a "semi-black box" he only inspects when problems arise.

Trend Insight: A Paradigm Shift from "Tool Use" to "Team Collaboration"

Willison's dilemma reveals a deeper trend: AI coding tools are evolving from "strictly reviewed tools" to "default-trusted virtual team members." This is not just an efficiency gain but a migration in work relationships and trust models. In a large company, you don't read every line of code of a service you depend on from another team; you rely on their documentation, API behavior, and track record. Now, developers are building similar, performance-based trust with AI agents.

But here lies a fundamental contradiction: human teams have "professional reputation" and accountability mechanisms; AI does not. An AI won't suffer reputational damage for writing bad code. Therefore, this trust is built solely on the AI's consistent and reliable performance. If the AI makes a serious error in a non-obvious place (e.g., a security flaw) and the developer, trusting it, fails to review, the responsibility falls entirely on the human engineer. This is the source of Willison's "unease" and "guilt."

Practical Value and Counter-Intuitive Insights

For developers in the AI era, this article's value lies in posing a practical question we must confront: How do we redefine "responsible development"?

1. Adjusting Review Strategies: Future code review may no longer be line-by-line but based on "functional modules" or "trust tiers." For clear, routine code generated by AI, review can be coarser; for core business logic or security-critical code, high-granularity review remains essential.
2. Building New Quality Assurance Systems: When the "author" shifts from human to a human-AI hybrid, the importance of testing, monitoring, contract testing, and formal verification will skyrocket. These automated quality gates will become the crucial defense to compensate for "reduced review."
3. Redefining the Engineer's Value: The core value of an engineer is shifting from "writing every line of code" to "defining problems, designing architecture, setting quality standards, and making final trust decisions." Experience and judgment are more critical than ever.

A counter-intuitive insight is that the ultimate goal of AI tools might not be to make programming "easier," but to make building "complex and reliable" systems "more accessible." Willison's ideal is to "build higher-quality stuff faster," not "build lower-quality stuff faster." When AI can reliably handle routine implementation, engineers' energy is truly freed for the high-value tasks requiring human creativity, trade-offs, and deep thinking. This quiet paradigm shift is redefining what "professional" software development means.