Jump Trading entrusts research to GPT-6 under human oversight

The quantitative market team deploys agents based on GPT-6 Astra to conduct long and iterative analyses, from coding to quantitative studies. Outputs remain subject to critical reviews and controlled execution. Lucas Baker describes a rapid progression towards more autonomous research, without taking the final word away from the teams.

Human supervision frames every signal produced by agents

Jump Trading operates in a highly regulated industry, where any mistake can lead to financial impacts and compliance issues. According to the company, the architecture of its system relies on precise limits, explicit constraints, and an infrastructure focused on supervision and transparency, complemented by human verification of changes. Lucas Baker believes that trust in the results provided by the agents is based on a secure environment, ultimately validated by human assessment. When an agent issues a trading signal, it is treated as potentially erroneous and combined with other signals within a rigorously controlled execution environment. These signals are integrated with other flows and are not executed without this control framework.

Long and iterative workflows, managed like with a colleague

Teams now approach AI as a colleague with whom they define a central problem, a working environment, and an evaluation method. Researchers guide one or more agents in real-time on where to focus the analysis and on the next task. Over the course of a long task, the system examines its own results, compares them against agreed-upon criteria from an initial proposal, and redirects its efforts without requiring human analysis at every step. Work can be configured to run for days, drawing from numerous data sources and making subtle judgments about the relevance of signals. The results are then explicitly evaluated against the defined criteria.

Current capabilities attributed to GPT-6 Astra and scope of use

Lucas Baker leads LLM R&D and agent development at Jump, which designs agents, infrastructures, and tools to deepen researchers' ideas. He attributes the scaling and complexity of reliable tasks assigned to agents to the GPT-6 series, particularly GPT-6 Astra, ranging from daily coding to advanced quantitative studies. He describes a new level of autonomy for long-term tasks requiring flexible coordination of agents and persistence on complex flows. According to him, the need for human intervention decreases when the environment is secure, well-monitored, and the objectives are clear. These practices apply to all time horizons and asset classes that the company deals with.

From code snippets to codebases and beyond

Over the course of a year, the use of AI has evolved from assistance in writing a code snippet to a system capable of building entire codebases and services. Agents are described as being able to identify significant changes and stack them in a recursive improvement process. Year over year, consistent progress in benchmark measures is accompanied by shifts in the models that can be executed, often difficult to anticipate a few months in advance. In 2024, having the first agents write a single error-free file was considered impressive. By 2025, creating complete codebases from scratch became feasible. In 2026, dynamic collaboration among multiple agents advanced certain unresolved research issues.

Towards a fleet of agents for supervised autonomous research

Lucas Baker anticipates that autonomous research, understood as the recursive improvement of measurable systems by research agents, will become common practice. At this stage, extended work involving GPT-6 Astra still includes regular checkpoints with the person defining the task, whether it involves choosing the data to extract, execution duration, importance criteria, or consistency of intermediate results. The advanced version described would always begin with human framing specifying inputs, environment, metrics, trade-offs, and priorities; the rest of the process would be delegated to a fleet of agents coordinated by other agents. In a well-structured pipeline, they would decide on exploration, allocate computation, and integrate promising results, aiming to derive a useful outcome from an open question. Baker posits that the ability to solve a Millennium problem would go hand in hand with discovering interesting facts in quantitative finance and wonders about future advancements.

Reminder of the core business and inherent limits of prediction

Jump Trading designs predictive models by cross-referencing market data, news, events, and alternative data sources. The complexity and noise of the markets make accurate forecasts rare. Lucas Baker emphasizes, however, that predictions that are only slightly better than chance, when applied at scale, can be sufficient to support a strategy.