harness matrix evals

support running our benchmark suite across deepagents + other harnesses ideally these would be regularly run so we can identify areas we need to improve and learn from other open harnesses

accordingly, we should have an up to date benchmarks page on our docs