01How Simon Willison Uses Coding Agents Safely and Effectively
Simon Willison describes his AI-assisted development workflow, emphasizing that coding agents become truly useful when paired with strong tests, clear repo conventions, and sandboxing. He argues that red-green TDD and automated/manual API checks let agents generate reliable code, while template-driven project scaffolding helps them follow existing patterns. The discussion also covers prompt injection, the 'lethal trifecta' risk model, and why isolating agents in containers or other sandboxes is crucial when they have access to private data or external communication.
Willison now often writes code on his phone and uses coding agents for rapid experimentation and feature shipping.
He says the major shift came when newer models started producing good solutions rather than janky drafts.
Red-green test-driven development is his default agent workflow, and he considers tests effectively free now.
He supplements tests with runtime checks like starting servers and calling APIs via curl.
He uses Showboat to document manual test runs in markdown.
Conformance test suites are especially powerful because they can drive implementation across multiple languages/frameworks.
Code quality still matters for long-lived codebases, but poor output is ultimately a choice if not reviewed and refined.
Existing project templates and consistent patterns help agents generate better code.
Prompt injection and the 'lethal trifecta' are highlighted as major security risks for LLM-based systems.
Sandboxing and limiting external communication are presented as the main defenses, especially for coding agents with access to sensitive data.