Hospital-bench from Zavoai tests agent reliability in hospital operations
The most instructive signal is that safety failures were unevenly distributed: models with similar accuracy scores differed sharply in harmful actions like leaking patient records or altering paid claims. Teams evaluating agents for regulated environments such as healthcare or finance should weight harm rates as heavily as task scores. Hospital-bench also gives developers a concrete, tool-rich simulation for stress-testing agents across 131 tools before real deployments.