Jump Trading, a quantitative trading firm, builds predictive models from market data, news, events, and a range of alternative data sources. Markets are complex, noisy, and change over time, so the goal is not certainty but a small statistical edge: per Lucas Baker, Head of LLM R&D, predicting even slightly better than a coin flip at scale is enough to make a strategy successful. Baker’s team builds the agents, harnesses, and infrastructure that let quant researchers explore ideas with greater breadth and depth, and he credits the GPT-6 series—especially GPT-6 Astra—with unlocking a new tier of autonomy for long-horizon tasks requiring flexible agent coordination and extreme persistence on complex workflows.

Over the past year, Baker says, AI has shifted from a tool for writing one-off code snippets and finding small bugs into a system that can develop entire codebases and services by itself. His team now treats AI more like a colleague: a researcher defines a key problem, a work environment, and a way of evaluating the quality and significance of results, then steers one or many agents in real time about where to focus the analysis or which job to run next. With GPT-6 Astra, agents can find meaningful and practical changes and merge and stack those wins through recursive improvement. During a single long-running task, the system analyzes its findings, judges them against the criteria from the initial proposal, and actively redirects its own efforts, so a person does not have to review each round of changes. That allows defining work that runs for days, pulls from many data sources, makes subtle calls about what is important, and produces a comprehensive analysis that actually works.

Because Jump operates across every time horizon and asset class in a heavily regulated industry, mistakes can have both financial and compliance consequences. Baker says it is critical to be aware of the risks of entrusting work to AI, and also that agentic intelligence can be applied to quality, security, and monitoring, not just feature work. The safety model relies on strong system design and boundaries, clear constraints, infrastructure that promotes steerability and observability, and human review of changes. As he puts it, a safe environment where the agent is free to produce any output, combined with a human acceptance process at the end where critical validation takes place, is what gives the team confidence. An agent-produced trading signal, for example, is scoped and reviewed like any other output—usually informative but potentially wrong—and integrated with every other signal in a stringently reviewed and controlled execution environment.

Baker predicts a future where “autoresearch,” the recursive improvement of measurable systems by agent researchers, becomes so ubiquitous it is simply part of a quant researcher’s ordinary workflow. Today even GPT-6 Astra‘s longest-running work involves regular check-ins with the person defining the task: what data to pull, how long to run, what counts as important, and whether intermediate results make sense. Advanced autoresearch would still begin with a human-defined system of inputs, environment definition, evaluation metrics, trade-offs, and priorities, but would entrust the rest to a loosely structured fleet of agents coordinated by other agents. Within a well-structured pipeline, those agents would decide how to explore ideas, allocate compute time, and integrate promising findings, starting from little more than an open question. Baker frames the ambition directly: ask a general question you do not know the answer to and have a useful result come back. He also notes the step changes in capability—writing a single file without mistakes was impressive for agents in 2024, creating entire codebases from scratch became possible in 2025, and making progress on open research questions with many agents collaborating dynamically arrived in 2026—and asks what the next year will bring.