Building the future of software requires moving beyond incremental plugins to reimagine the development environment as an AI-native ecosystem.
The Fallacy of the Incremental Extension
When GitHub Copilot first emerged, the tech world viewed it as a finished revolution. To many, it seemed like the 'Goliath' had already won the space, leaving only crumbs for startups. However, this view missed a fundamental distinction: the difference between an extension and an environment. Most early AI coding tools were built as plugins—incremental improvements designed to fit into existing workflows. They were helpful, but they weren't transformative. They didn't change the nature of the craft; they just made the typing faster.
The real opportunity lies in the total automation of the software development lifecycle. Within the next five years, every stage of building software—from architecture to debugging—will flow through AI models. To capture this, one cannot simply build a better 'autocomplete.' You have to build a tool that is AI-native from the ground up, a system that understands the entire codebase as a cohesive unit rather than a series of isolated text files. This requires moving past the 'armchair MBA' logic of finding a quiet, uncompetitive niche and instead tackling the hardest problem in the most crowded room.
Lessons from the Desert
The path to building a category-defining tool is rarely linear. Many founders fall into the trap of 'box-checking'—pursuing ideas because they look good on paper rather than because they solve a visceral problem. We spent months wandering the desert, attempting to build AI for mechanical engineers or encrypted messaging systems for journalists. These projects failed not because the technology was poor, but because we lacked the domain obsession required to see them through. We were trying to be clever about market gaps instead of being honest about our own interests.
When we finally pivoted to building a code editor, we faced immediate skepticism. Critics argued that either AI was a fad or that AGI would make all software applications obsolete within two years. We chose the middle road: the belief that AI would be transformative but would require years of building specialized, high-friction interfaces to be truly useful. This 'monk phase' of development—ignoring growth hacks to focus entirely on product quality—is where the real compounding happens. In a world of 'movie magic' demos, the only thing that scales is a tool that actually works when the user is alone at 2:00 AM.
The Chromium Strategy and Model Muscle
One of the most difficult decisions for a developer-tool startup is deciding what to build from scratch. Initially, we spent months building a code editor from the ground up using open-source primitives. We quickly realized that competing with twelve years of VS Code development was a distraction. To win, we had to adopt a 'Chromium-style' strategy: use the industry-standard engine as the foundation so we could focus exclusively on the AI layer. This allowed us to stop worrying about file explorers and focus on 'next edit prediction' and codebase awareness.
Pragmatism must also extend to the models themselves. While there is a romantic allure to training a massive custom model from day one, the cost and effort can be a death trap for a lean team. We started with off-the-shelf models to find product-market fit, only building 'model muscle' when we needed to push past the limitations of standard APIs. By training smaller, specialized models for specific features like speed and accuracy in multi-file edits, we could outperform general-purpose giants. A lean team of fewer than ten people can outmaneuver a conglomerate if they are disciplined about where they apply their technical leverage.
The Future of the Craft
We are entering what I call the 'long messy middle' of software development. AI is beginning to act like an advanced compiler, hiding layers of complexity from the developer just as C++ once hid the complexity of assembly language. However, this does not mean the end of the engineer. Professional developers will still be required to read, review, and orchestrate the logic. The skill set is shifting from rote syntax to high-level system design and rigorous verification.
For the next generation of engineers, the advice remains the same: ignore the institutional pressure to check boxes. The most valuable skill is the ability to build something substantial over a long period, driven by genuine curiosity rather than a resume requirement. Programming is becoming a foundational 'general education' skill, much like mathematics. Even as the tools change, the ability to think logically and solve problems within a complex system will remain the most valuable currency in the modern economy. The goal isn't just to write code faster; it's to change how we think about building the future.