Descriptions:
At the Pragmatic Summit, Eric from Stedig interviews Simon Willison—co-creator of Django, creator of Datasette, and maintainer of hundreds of open-source repositories—about the engineering practices that make coding agents reliable in production. Willison opens by demonstrating his phone-based development workflow live: minutes before the interview he prompted Claude Opus 4.6 to optimize a Python-based WebAssembly engine and received a 45-49% speed improvement on a Fibonacci benchmark with a single instruction.
The conversation maps the stages of AI coding adoption—from asking questions, to AI writing code, to AI writing all code, to the emerging frontier where no one on a team reads code either—and examines what engineering discipline the final stage demands. Willison explains his use of cookie-cutter project templates that enforce consistent file structure, testing patterns, and CI configuration before an agent touches the codebase, arguing that agents reliably follow established patterns the same way junior developers copy-paste from existing code. High-quality codebases produce higher-quality agent output.
Willison also discusses prompt injection at length—a term he coined roughly three and a half years ago—explaining why LLMs are ‘incredibly gullible by design’ and how the lethal trifecta of autonomous action, external data access, and insufficient output validation creates exploitable attack surfaces in agent systems. He references StrongDM’s controversial ‘software factory’ model, where nobody writes or reads code, and explains the testing and verification rigor required to make such an approach anything other than ‘wildly irresponsible’—particularly for security software. The interview is essential viewing for engineers building or evaluating coding agent workflows.
📺 Source: The Pragmatic Engineer · Published March 19, 2026
🏷️ Format: Interview
