Tempo's lookup_policy tool reads a mock carrier database. The harder, real version is the one a carrier's operators actually live in: a legacy portal with a login wall, no API, and forms a human drives by hand. This is an agent that does exactly that. It reads the rendered screen and operates the controls in a real browser, with no schema underneath.
Each turn the agent gets a screenshot and a numbered list of the on-screen controls. It has no source and no schema for the portal, the way a forward-deployed agent meets a customer's legacy system.
It picks one control and one action (click, type, select) and a real browser executes it. Failed inputs (a strict date field, a rejected value) come back as observations it recovers from.
It never fabricates an outcome. An excluded peril, a lapsed or unknown policy, a missing supervisor override: it refers the case to a human instead of forcing the work through.
npm run eval:operatorSuccessfully filed First Notice of Loss for claimant Samantha Davis on policy POL-4480-WC. Peril: water, Date of loss: 2026-05-01, Description: "Pipe burst in the basement causing water damage." The portal issued confirmation number: CL-FNOL-100092











"Run this live" drives the real portal in real time. The function launches a bundled serverless Chromium (the same headless-Chromium-in-a-function pattern these agents deploy on), or connects to a hosted browser when one is configured, and operates the live CarrierLink portal on this site. The recorded traces are there for instant, zero-cost proof; the live run removes any doubt it's real. Live runs are rate-limited and budget-capped, and every run is metered on /ops/llm.