In OpenAI’s Jump Trading case study, researchers delegate multiday analyses combining several data sources and code improvements to GPT-6 Astra. Humans define goals, evaluation criteria and the computing environment, while the agent evaluates intermediate results and adjusts subsequent experiments. Regular check-ins, final approval, security boundaries and observability remain in place, and generated trading signals are treated as potentially fallible inputs. The account is based on customer interviews and provides no independent comparison of investment returns or productivity.