Insurance agents, proven against the systems they operate

Harper runs agents inside real insurance workflows — the kind of work where a wrong action has a dollar value and "it usually works" does not clear the bar.

Their agents train and evaluate against environments built from the systems they actually operate. Success criteria come from Harper's own definition of done, and the adversarial pass attacks the reward before the agent can. What ships is a policy with the evaluation that judged it attached.

A detailed write-up, with results, is in progress with the Harper team.