Multi-model benchmarks

One challenge. Several agents. One reproducible environment. See which one actually works. Comparison is a core product, not an add-on.