When AI Writes Most of the Code, Software Engineering Changes

Summary

The article argues that AI coding agents have reached a capability threshold where they can generate most of the code for many low-risk, well-tested projects. The author describes using Opus 4.5 and GPT-5.2 to produce production code through prompts, reviews, and tests, and using Claude Code from a phone to create pull requests and run changes through GitHub Actions. Reports from engineers including Jaana Dogan, Malte Ubl, David Heinemeier Hansson, Andrej Karpathy, and Boris Cherny reinforce the sense that late-2025 model releases changed expectations; Cherny said Opus 4.5 wrote every line in roughly 200 pull requests over a month. The article treats Anthropic CEO Dario Amodei’s prediction that AI would write 90% or more of code as increasingly plausible, while noting that the evidence is partly self-reported and most convincing for greenfield, low-risk work with strong tests. It predicts that prototyping, language-specific expertise, routine tickets, and refactoring will become less valuable because agents can handle them across stacks. In contrast, engineers who can define requirements, specify nonfunctional constraints, design architecture, build testing and observability loops, manage technical debt, and ensure security, performance, reliability, and scale may become more valuable. More generated code could also mean more bugs, security issues, resource use, and review overload; the article cites a 30% increase in deployment change-failure rates in a Cortex survey of more than 50 engineering leaders. Poor engineering practices may therefore cause harm faster, while mobile and browser-based workflows could weaken work-life boundaries. Product managers and engineers may overlap more as both prototype and direct agent work, with smaller teams and fewer formal handoffs. The author concludes that software engineering is unlikely to disappear, but its center of gravity will move from typing code toward deciding what to build, constraining and validating implementation, and remaining accountable for production systems.