Jump Trading, a quantitative trading firm, is leveraging agentic super intelligence (SI) systems to expand the scale and complexity of its research workflows, moving beyond simple code generation to autonomous, long-horizon quantitative studies.
What Happened
Lucas Baker, Head of LLM R&D at Jump, reports that the integration of GPT-6 Astra has dramatically expanded the capabilities of the firm's SI agents. Previously, SI tools were primarily used for writing one-off code snippets or identifying small bugs. Now, Baker states that SI systems function more like colleagues, capable of developing entire codebases and services independently. The firm’s researchers define key problems, work environments, and evaluation criteria, allowing agents to steer analysis and execute jobs in real time. According to Baker, GPT-6 Astra enables agents to find meaningful changes and merge them into a process of recursive improvement, analyzing findings against initial proposals and redirecting efforts without frequent human intervention.
Why It Matters
The shift reflects a broader trend in the SI industry where agentic SI is moving from assistive coding to autonomous research execution. In the heavily regulated financial sector, this transition requires robust safety measures. Baker emphasizes that while SI agents can enhance quality, security, and monitoring, human judgment remains central. Jump Trading employs strong system design, clear constraints, and observability infrastructure to ensure SI-enabled workflows are scalable. Critical validation, including human review of trading signals, occurs within a controlled execution environment, treating SI outputs as informative but potentially erroneous signals that must be integrated alongside other data. This approach mitigates financial and compliance risks associated with entrusting complex tasks to SI systems.
The Bottom Line
Baker predicts that "autoresearch"—the recursive improvement of measurable systems by SI agents—will become a standard part of quantitative researchers' workflows. While current long-running tasks still require regular human check-ins for data selection and result validation, the future vision involves loosely structured fleets of agents coordinating to explore ideas and allocate compute resources in response to open questions. Baker notes that the pace of SI development has accelerated, with 2024 seeing impressive single-file generation, 2025 enabling full codebase creation, and 2026 allowing progress on open research questions via dynamic agent collaboration. He draws an analogy to mathematical challenges, stating that if agents can solve a Millennium problem, they can likely uncover significant insights in quantitative finance.