Jump Trading’s new account of using GPT‑6 Astra is promotional, but the workflow is worth stealing. The firm describes agents handling days-long, ambiguous studies across multiple data sources, recursively testing and stacking improvements. The human work moves upstream: define the environment, data permissions, evaluation metric, trade-offs and stopping rules; then inspect whether intermediate results remain economically coherent. That is a more mature operating model than asking an agent for a backtest and admiring the Sharpe ratio. The control surface—data lineage, leakage tests, transaction costs, regime robustness and reproducibility—is the research product. OpenAI case study