recursive-benchmark

You can benchmark coding agent performance by running paired benchmark arms that compare recursive-mode off versus on. The framework generates disposable repos with logs, timings, and a final markdown report for auditable results.

The best way to compare recursive-mode performance is through a structured, repeatable framework that automates the generation of two comparable benchmark arms from a single project spec, ensuring reproducible results and auditable evidence.

Yes, you can generate disposable benchmark repos with logs for multi-arm evaluations. The process creates isolated environments, bootstraps recursive scaffolds in the on-arm, and captures comprehensive logs and artifacts for review.

No external dependencies are required to run paired evaluations. The framework operates independently to bootstrap recursive scaffolds, apply run-local requirements, and produce a final comparison score without needing additional tools.

To produce a final markdown report after running recursive-mode benchmark tests, the framework automatically captures logs, timings, and artifacts from both the off and on arms, compiling them into a comprehensive comparison document for audit.

You should not use a paired benchmark approach if your project lacks a clear single spec for generating comparable arms, as the framework relies on isolating recursive-mode off versus on from the same project requirements for fair evaluation.