Simon Willison Says AI Coding Agents Finally Hit Their Inflection Point

Simon Willison, one of the most respected voices in the developer community, believes AI coding tools have crossed a critical threshold. In a wide-ranging new interview with Lenny's Newsletter, Willison argues that the November 2025 releases of GPT-5.2 and Opus 4.5 represent a genuine inflection point for artificial intelligence agents that write code. His observations carry weight because he has spent decades building developer tools and closely tracking how large language models reshape the craft of software engineering.

What Changed in November 2025

Willison says the latest generation of large language models, often called LLMs, finally made agentic coding reliable enough for daily professional use. LLMs are the artificial intelligence systems that power tools like ChatGPT and Claude, generating text and code from written prompts. Unlike earlier versions that produced inconsistent and sometimes incorrect results, these newer models can now handle complex multi-step programming tasks with far fewer errors. Willison has adapted so completely that he now writes roughly 95 percent of his code from his mobile phone, relying on AI agents to handle the heavy lifting while he reviews and guides their output.

The Patterns That Make It Work and the Risks Ahead

Willison identified three key agentic engineering patterns driving this shift. The first is red and green test-driven development, where developers write tests before code and let AI fill in the implementation until all tests pass. The second uses templates as scaffolding to give AI agents structured starting points for common tasks. The third involves hoarding reusable code patterns that AI tools can reference across multiple projects to maintain consistency. However, Willison warns that prompt injection remains a dangerous and fundamentally unsolved problem. He calls it the lethal trifecta, a scenario where private data exposure, untrusted content, and external communication channels combine to create serious security vulnerabilities. This means AI agents that access sensitive information could still be manipulated by attackers in ways that are difficult to detect or prevent.

Despite the clear productivity gains, Willison admits the cognitive demands of working alongside AI are surprisingly intense. He reports feeling mentally exhausted by 11 in the morning most days, suggesting that managing AI coding agents requires a fundamentally different kind of sustained focus compared to traditional software engineering.