Quality anyone?

Cursor announced the release of GPT 5.2 Codex, a new frontier model designed for long-running tasks. The CEO Michael Truel revealed that they used this model to build a browser entirely with AI, running uninterrupted for one week. The AI wrote over 3 million lines of code across thousands of files and created a rendering engine from scratch that handles HTML parsing, CSS cascade, and other browser fundamentals. However, there's one critical caveat: the browser only 'kind of works,' which becomes the central theme of the video's argument about software quality and AI development.

The speaker argues that while it's impressive AI built a working browser, the real complexity in any project lies in the final 20% that takes it from 80% to 100% completion. This principle applies not just to browsers but to virtually any project in life. For many software products, achieving only 70-80% functionality isn't sufficient, and that extra bit can be extraordinarily difficult to achieve. The speaker suggests that AI alone might not be capable of delivering that crucial final quality level. This sets up the core tension between impressive AI capabilities and the reality of production-ready software requirements.

The speaker emphasizes being 100% positive on AI and uses it extensively in their own work. They cite buildmygraphic.com as an example, where most code was written by AI, but crucially with human instructions, code reviews, and manual tweaking when necessary. They recently released a major update to their 'AI for Developers' course covering GitHub Copilot and Cursor features. The speaker firmly believes AI is the future for developers and that using it heavily and efficiently will be vital. This section establishes that the critique is not about AI usage itself, but about how developers approach AI integration in their workflow.

The speaker introduces a spectrum between vibe coding and agentic engineering. Vibe coding is defined as letting AI write code with no code reviews, no understanding of the codebase, and no specific instructions about patterns or packages—essentially zero developer knowledge or involvement. The speaker argues this pure form of vibe coding does not have a future for commercial or real products. However, they acknowledge different definitions exist and that their definition represents the most extreme form. The key point is that developers should position themselves somewhere on this spectrum of AI usage rather than avoiding AI entirely.

The speaker identifies legitimate scenarios where pure vibe coding is viable: personal utility tools, throwaway software used once or twice, and potentially free software where quality standards are lower because users aren't paying. In these contexts, it doesn't matter if the code covers all edge cases or has potential bugs, as long as it gets the immediate job done. For example, requesting a quick script to accomplish a one-time task is perfectly appropriate for vibe coding. This nuanced perspective shows the speaker isn't dismissing vibe coding entirely, but rather defining appropriate contexts for different approaches to AI-assisted development.

Agentic engineering is presented as the future approach, where AI is used as a tool rather than autonomous creator. This methodology includes giving AI clear instructions about patterns and libraries, reviewing code regularly (potentially with automated AI-assisted reviews), and having developers directly engage with code when AI gets stuck or when establishing initial implementations. The speaker emphasizes that this doesn't mean AI only handles trivial tasks—complex tasks are absolutely included—but rather that human oversight and architecture decisions remain central. The speaker acknowledges they could be wrong and pure vibe coding might become viable in a few years if AI improves dramatically, but currently believes agentic engineering is the right path.

The speaker analyzes Cursor's motivation for this announcement, noting that recent weeks have been dominated by developers using Claude Code with the ROLF loop for vibe coding-style autonomous development. Cursor's demo appears to be an effort to shift the narrative and show they can also handle long-running autonomous tasks. The speaker acknowledges Cursor is an amazing tool and understands the competitive positioning, but expresses concern about the 'kind of works' attitude. They emphasize the only wrong decision is not using AI at all—developers should be somewhere on the spectrum between vibe coding and agentic engineering.

The speaker identifies a broader software quality crisis that predates AI: operating systems like iOS and Windows becoming buggier, video games being unplayable on launch day, and general software quality degradation over 15+ years. This decline stems from an 'iterate quickly and patch later' mindset that developed in the industry. The speaker warns that AI is accelerating this trend because it enables even faster patching and may lead developers to care less about bugs (since they can be fixed instantly) or code quality (since humans don't need to read it). Vibe coding could result in codebases full of messy if-statement patches rather than clean implementations, representing a dystopian future for software quality.

The speaker concludes with an optimistic vision: there will be a genuine market for high-quality software that isn't broken on release. There's no requirement to sacrifice software quality for speed—AI can be used to build better software by combining human skills with AI capabilities. Developers can use AI as an extra pair of eyes for code review and quality assurance. While the speaker expects most companies will continue prioritizing speed, they believe valuable opportunities will emerge for companies and developers who use AI to achieve the best of both worlds: high quality and efficient development. This represents a call to action for developers to resist the race-to-the-bottom in software quality and instead leverage AI for excellence.