Research
N-Version Programming with Coding Agents
Tests whether running multiple coding agents and voting improves reliability, using 48 agent-generated implementations of the Knight-Leveson Launch Interceptor spec against 1,000,000 random inputs. Majority voting over three-version units drops mean failures from 387.44 (single) to 130.99 (triples), with 11,000+ N-version units showing zero observed failures — though common-mode failures persist in ambiguous spec areas. Quantifies the payoff of ensembling diverse coding agents.
Source
↳ Follow the thread