Measuring agent reliability in the wild

Lab benchmarks rarely predict how an agent behaves on a Tuesday afternoon with messy data. We track completion, correction and escalation rates across real workspaces instead.