Ensuring your code works when AI testing isn’t enough | BRK208
Simon Willison discusses how to keep software reliable when AI-generated code and AI-driven testing fall short, focusing on practical engineering habits that scale to both humans and machines.
Overview
This Microsoft Build 2026 breakout (BRK208) focuses on building AI-assisted software systems that still behave predictably under real-world constraints.
Key themes called out in the session description and chapter list include:
- How teams should think about productivity measurement when AI is generating code, and the resulting quality concerns.
- “Active refactoring” as an approach to keep codebases maintainable as AI assistance increases throughput.
- Guidance on documentation style, including avoiding opinionated or promotional tone.
- Using sandboxed iframes and their secure isolation properties as a concrete example of building safer, more reliable systems.
- “Agentic engineering” as a framing for professional AI-assisted coding, including parallels to other long-running engineering improvements.
- A discussion referencing The Mythical Man-Month and how its lessons map to modern AI/agentic development.
- Language/tooling considerations (e.g., Go vs Rust) in the context of readability, testing, and model understanding.
- Using AI to work in unknown languages and to build native apps via agentic coding workflows.