A twelve-hour coding-agent experiment creates a larger review problem

Simon Willison describes asking Claude Code to build a Rust extension that runs SPARQL graph queries inside SQLite. The agent worked for more than twelve hours and produced a large project with many packages. Willison says he cannot feasibly review the generated code and is unsure how valuable it is without investing time to understand it. The experiment exposes a gap between generating software and being ready to maintain it.

Simon Willison · Oct 16, 2026

Why it matters

The cost of understanding and owning generated code can remain substantial after the agent finishes its work.

More context

This is a personal experiment, not a production-readiness endorsement. The original post links the generated research directory and describes requesting documentation for each package plus a tutorial. Watch the experiment for workflow lessons; do not adopt the extension on the strength of its generated feature or speed claims.

Keep in mind

The implementation’s claimed completeness and benchmarks were not independently verified. This card covers the author’s reported process and review constraint, not a validation of SPARQL correctness or performance.

Original read