Why is AI now able to write complex code? Simon Willison's analysis

AI Model Breakthrough in 2025

• 💡 In 2025, AI models experienced a qualitative leap in capability, moving beyond simple code correction to handling overall software architectural design.
• 📈 Benchmarks like GPT-Dual showed significant performance jumps, with models like GPT-5.2 achieving 70.9% compared to older models' 38.8%.
• 🧠 This advancement was driven by a vastly expanded context window and enhanced reasoning abilities, enabling AI to tackle complex, multi-hour tasks and resource leaks.

Emergence of Coding Agents

• 🚀 A crucial development was the introduction of coding agents, which are autonomous systems capable of writing, executing, testing, and self-correcting code.
• ✅ These agents function as active development partners, not just tools, by independently looping through the development cycle from instruction to correction.
• 📊 Simon Willison's project demonstrated an AI agent writing 2000 lines of code for a single task, and generating 259 pull requests in a year, signifying a massive scale-up in productivity.

The Evolving Role of Human Developers

• 🤝 Human developers' role is transforming into "vibe engineering," where they guide AI with abstract goals and "feelings" rather than precise, detailed instructions.
• 🎯 Managing AI is now primarily a management problem, requiring skills to effectively direct and maximize the potential of highly capable AI "subordinates."
• 🔑 New essential skills include setting clear goals, providing necessary context, and offering evaluative feedback to AI outputs, leveraging human experience to amplify AI's power.

A New Era of Software Development

• ✨ The revolution in coding is attributed to the synergistic combination of intelligent AI models (the brain), autonomous coding agents (the limbs), and human guidance (the commander).
• 🗓️ Simon Willison concludes that 2025 was the "year of coding agents," marking the dawn of a new era where AI acts as a true partner in software development.