Researchers have created a sandbox called Thinkingbox to test the reliability of artificial agents in complex business workflows. The sandbox allows for isolated tool sessions and evaluates outcome based on terminal backend state. A benchmark, Thinkingbox-bench, contains 507 policy-conditioned workflows across various industries, including retail and insurance. Results show that even top-performing models struggle to complete tasks reliably, with a large gap between occasional success and consistent completion. The authors release both the sandbox and benchmark for further research. --- Why it matters: This matters because it highlights the limitations of current AI evaluation methods, which often focus on individual task performance rather than end-to-end workflow completion. Understanding these challenges is crucial for developing more reliable and robust AI systems that can handle complex business tasks. Source: https://arxiv.org/abs/2608.19741