Independent benchmark tests six coding agents across seven local AI models

A developer conducted an extensive benchmark test evaluating six AI coding agents across seven locally-run models. Each agent was run thirty times per model, with performance measured by task completion rates. The test revealed significant performance differences among agents, particularly when models generated tool calls in non-standard formats. The full methodology, raw results, and analysis are publicly available on GitHub. The developer disclosed they created one of the tested agents, Polyglot, and made all data public for transparency.