Daniel Azuma, a former Google engineer, publicly recanted his long-standing skepticism of AI coding tools in a blog post published on November 12, 2026. After leaving Google in September 2025 amidst company changes driven by AI priorities, Azuma spent his career break rigorously testing tools like Claude Code. His conclusion: the technology has fundamentally altered software engineering, succeeding at novel tasks with a quality he previously deemed impossible for large language models.

From Autocomplete to Reasoning Engine

Azuma’s initial dismissal of LLMs was rooted in the belief that they were merely 'glorified autocomplete' incapable of true reasoning or handling tasks absent from their training data. He expected to prove that AI could only regurgitate common, routine code. However, his experiments with the Ruby command-line framework Toys shattered this assumption. Claude Code correctly utilized project-specific terminology such as 'mixins' and 'flags,' demonstrating an understanding of the codebase’s unique DSL and internal modules that rivaled mid-level human engineers. The skeptic’s conversion accelerated when he tasked the AI with migrating a static website from 1990s-era PHP to Jekyll. The model reverse-engineered the legacy layout system and reproduced the content accurately, even self-correcting a hallucination when prompted to check its work. Azuma further tested the tool’s limits by assigning it a delicate concurrency fix in a complex library and a ground-up implementation of a Ruby client for the Hermes agent gateway API. In the latter case, the AI wrote its own probing tools to query the live API, effectively debugging the documentation against reality before writing the client code.

The Human Element in Engineering

Despite his enthusiasm, Azuma distinguishes between 'coding' and 'engineering,' arguing that the latter remains irreducibly human. He highlights that AI tools lack the ability to inhabit the physical world, meaning they cannot fully grasp the social, ethical, and aesthetic contexts required for true engineering accountability. While models can handle logic and syntax, they struggle with the 'taste' required to navigate how a system interacts with human communities and real-world constraints. This distinction necessitates a new set of skills for developers, including rigorous context management and the creation of meta-documentation like AGENTS.md. Azuma notes that best practices such as test-driven development and clean architecture are now more critical than ever, serving as guardrails for hallucination-prone agents. He also cites research by Mike Moore, which found that AI agents often misinterpret communication signals from neurodiverse individuals due to biases baked into their reinforcement training, reinforcing the need for human oversight in team dynamics.

Key Takeaways

  • AI coding tools have surpassed skepticism by successfully handling novel, complex tasks like legacy migrations and API reverse-engineering.
  • The role of the software engineer is shifting from writing code to managing AI agents, providing context, and enforcing architectural best practices.
  • Ethical and social understanding remains a human-only capability, making 'engineering taste' a critical differentiator in the AI era.

The Bottom Line

Azuma’s confession validates what many builders suspected: the 'slop' argument is a user error, not a tool failure. If you are still treating AI as a simple autocomplete rather than a pair-programming partner, you are falling behind. The future of software engineering is not about competing with AI on speed, but about leveraging it to handle the mechanical aspects of coding so humans can focus on systems design and ethical alignment.