Meta's Agent Problem Is Not Benchmarks

Meta's reported Watermelon benchmark claim is interesting, but the harder question is whether Meta can turn model capability into agentic products people actually trust and use.