AI Coding Moves Beyond Speed, Emphasizes Quality and Deliberation

The AI Coding Shift: From Speed to Deliberate Quality

The dominant narrative around AI-assisted software development has been one of unbridled speed: generating vast amounts of code with minimal human oversight. However, a significant counter-movement is gaining traction among developers. This approach leverages AI not as a 'slop cannon,' but as a meticulous partner for improving code quality, uncovering deep-seated bugs, and fostering a more thoughtful development process.

In a detailed blog post, software engineer Nolan Lawson argues that large language models (LLMs) are exceptionally flexible tools. While they can spew out code rapidly, they are equally potent when used to write high-quality code more slowly. This philosophy directly challenges the 'vibe-coding' trend, where AI tools are prompted to generate code using natural language, often leading to a deskilling of the development process.

AI as a Multi-Model Bug Hunter

The core of this deliberate approach lies in using AI for rigorous, multi-agent code review. Lawson details a custom workflow inspired by techniques from Anthropic's Mythos project, which demonstrated that LLM agents are exceptionally adept at finding bugs. His method involves running multiple AI models—such as Claude sub-agents, OpenAI's Codex, and Cursor Bugbot—in parallel to review a pull request.

These agents are tasked with finding bugs ranked by criticality. The key insight, adapted from an external article, is that using diverse models reduces the risk of hallucinations and false positives. The human developer then synthesizes these findings, validates them, and produces a final report. This process routinely uncovers a wide range of issues, from critical security flaws to minor stylistic inconsistencies.

The real challenge is not finding bugs, but prioritizing and validating them. This workflow often reveals pre-existing bugs in the codebase, sending developers on 'side-quests' to write unit tests and fix subtle flaws. This methodical improvement of overall code health is positioned as a more satisfying and sustainable practice than chasing raw output metrics.

The Evolving AI Coding Tool Landscape

This shift in developer philosophy coincides with a dynamic battle in the underlying AI model market. According to industry reports, Anthropic's Claude Code is gaining significant ground within startups. Founders cite its 'agentic workflow' as superior for complex tasks that span multiple repositories or require architectural decisions.

While Cursor remains in use, it is increasingly described as a fading secondary tool, valuable for simpler tasks but less capable for demanding work. Nearly every line of code shipped by some startups is now AI-generated, then reviewed and refined by senior engineers. The gap between what AI can handle and the responsibilities of a senior engineer is rapidly shrinking.

Google is also making strategic moves to compete. Instead of immediately releasing its largest model, Gemini 3.5 Pro, the company has prioritized the smaller, faster, and cheaper Gemini 3.5 Flash to power its Antigravity AI coding service. The goal is to generate a massive, anonymous dataset of developer interactions—noting where engineers halt tasks due to poor AI output—to refine the larger Pro model through reinforcement learning.

The Human-in-the-Loop Paradox

This trend towards deliberate AI coding underscores a broader paradox identified in automation. Dan Shipper, CEO of Every, notes that as companies automate knowledge work with AI, the demand for human judgment to direct, review, and improve that automation actually increases. AI commoditizes yesterday's competence, inflating the value of expert oversight.

Shipper's internal benchmarks highlight this: an AI agent given a clear, principled instruction to perform a structural rewrite scores highly. When prompted merely to 'solve all the errors,' its performance collapses. The model's output is inseparable from the quality of the human-constructed frame around the task. This creates an 'abundance cycle' where cheaper skills see expanded demand, but the need for human discernment to elevate adequate output to good output becomes paramount.

A Philosophical Divide on Efficiency and Craft

This movement towards deliberate AI use exists in tension with a more skeptical view. Some commentators, like Wendy Liu, consciously avoid AI tools, viewing the struggle of thinking and problem-solving as integral to the human experience and professional craft. They see inefficiency as a cost worth paying to preserve intentionality and deep understanding.

This perspective reminisces about an earlier era of hands-on learning, where debugging and studying documentation were core to the craft. It stands in stark contrast to a landscape where high-school dropouts can leverage AI to build startups and slick apps can be spun up with minimal coding knowledge.

The Future: Supercharging Methodical Development

For proponents of the deliberate approach, AI becomes a force multiplier for a specific kind of developer: the careful, quality-obsessed engineer. Lawson suggests using AI not just to write code, but to explain how a PR works, how it might fail, and to generate documentation like Mermaid charts. Skills like Matt Pocock's '/grill-me' prompt can be used to gain deep understanding of complex changes.

The outcome may not be a dramatic increase in 'lines of code per day.' Developers might burn tokens only to discover a fundamental flaw in their approach. However, the result is a supercharged version of methodical programming: a process focused on making the codebase more robust, understandable, and maintainable for the next developer. In the race to automate coding, the most enduring impact of AI may be its ability to make us not just faster, but more thorough.