Does it actually work?

Every query in backend/tests/queries.json run through the live three-agent workflow, scored against what each one is expected to do.