How to use Claude to compare NURBS implementations

The author ran a comparison on an experimental NURBS surface modeling project, giving Astra and Opus 5.5 the same repository state, task packet, and test suite. Both were allowed to delegate implementation work to lesser sub-agents, provided they managed and reviewed the output. The author reports that Astra ran for 13 hours and Opus for 20 hours, but cautions that test-suite frequency largely determined elapsed time.

A faulty unit test became a dividing point. Opus interpreted the written requirements, identified the bug, and corrected the test. Astra initially tried to preserve the suite and ultimately abandoned an implementation that still stubbed critical B-spline functionality. The author integrated useful Astra tooling into the Opus version, then gave Astra a hint and permission to change the test. That improved its progress without producing a working implementation; Opus's version was also incomplete.

Fresh agents reviewed both repositories twice per model. Their initial preferences followed their own model family. The author collected the four reports for another review, after which both models preferred the Opus implementation as the basis for further work. Who it's for: developers comparing agent approaches to a demanding numerical implementation while retaining responsibility for requirements and test validity.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: Hacker News · rspeele