Jun 17, 2024
A benchmark evaluates multi-turn agents using simulated customers, business rules and API tools in retail and airline tasks.
Final database states measure task success; repeated trials test consistency. Code and data are shared under MIT, while the simulated setting has limits.