Introduction

You can download the PDF here.

The Illusion of Acceleration

While artificial intelligence has surpassed human performance in isolated programming tasks, such as translating well-scoped prompts into working code (Chen et al. 2021), it has not replaced the software engineering department. Over the past few years, a wave of tools has promised to enhance productivity: GitHub Copilot (GitHub 2021) suggests code as a developer types; Devin claims to function as an AI software engineer (Cognition AI 2024); and tools like Sweep.dev (Sweep.dev 2025), Continue.dev (Continue.dev 2021), and Magic.dev (Magic.dev 2024) push even further, embedding LLMs into developer workflows or forming agentic teams.

But across all of these systems, one assumption remains: human developers will stay in the loop. They are expected to review AI-generated code, fix edge cases, manage systems, and adapt software as business needs evolve. This assumption is not just limiting; it is the heart of the problem.