A 2026-2031 scenario for automated quant teams

Our central scenario is that a small team gains access to much of the routine work of a research desk before it gains the economics of a leading trading firm. Agents can lower the cost of producing models, maintaining software and monitoring a mandate. Deploying capital successfully still depends on the market, the information and the operating system around those agents.

We make three conditional projections: supervised task bundles in 2026-2028, bounded autonomous programs in 2028-2031 and broader organizational delegation from 2031 onward. These windows are DXRG assumptions as of October 2026. They are intended to be updated against observed acceptance results, with no claim that Jane Street or METR announced these dates.

The official role descriptions provide a useful constraint. Jane Street describes research and trading as overlapping activities. Its ML roles connect model invention with engineering and optimization. A credible future quant team needs those functions to work together, even if agents perform much of the work previously assigned to individual people.

2026-2028: a small team can run more of the loop

We expect the first integrated bundle to start with an explicit research question and end with a reviewable proposal. An agent prepares authorized data; another implements the experiment; a reviewer checks evaluation validity; an execution runtime applies a separate mandate. Those functions may use one model or several, but independence of critical checks matters more than the number of agents.

In this scenario, a two-person team could cover a narrower market universe with substantially less manual preparation than a traditional workflow. That is an illustrative organizational design, rather than a measured staffing result. Its advantage comes from reducing coordination and production cost, especially when information is public and decisions allow minutes of analysis.

Here, mid-horizon means forecasts and holding decisions over hours to days. We expect standard public-signal workflows in that range to face increasing commoditization: more participants can generate similar features, test similar hypotheses and obtain competent implementations. Open markets may see more experimentation while the rewards for conventional analysis compress.

Higher confidence attaches to cheaper research production and operational summaries. Confidence in durable excess returns is lower because competitors can copy methods and because trading costs change with adoption. The near-term acceptance requirement is a completed loop whose records allow an independent reviewer to reconstruct inputs, proposals and outcomes.

2028-2031: the team delegates a program

Our second projection is that agents can maintain a restricted research and trading program over many cycles. They allocate an approved compute budget, retire failed candidates, diagnose incidents and recommend changes within an explicit mandate. Human work shifts toward selecting market scope and evaluating the evidence for capital allocation.

The mechanism requires persistence plus dependable self-correction. A failure in one cycle must change the next investigation without contaminating the final evaluation. The METR original paper provides a conditional reason to expect longer software-task execution as capabilities improve. Translating that progress into an ongoing market program adds changing distributions and economic feedback, so the calendar remains our scenario.

A decisive evaluation would freeze an information universe and compare equal-budget programs across unseen time periods. Report predictive scores, operational failures and simulated after-cost outcomes separately. Then examine how a limited live deployment differs from the simulation under its authorized scope. Each stage answers a different question about the integrated team.

Our confidence increases when performance survives independent review, realistic execution and market changes. It decreases when the agent needs frequent hidden rescue or repeatedly rewrites the conditions that define success. Failure of those conditions would delay the second window while leaving the first window's productivity gains intact.

2031 onward: more autonomy, persistent competitive differences

Our longer-range scenario allows agents to operate increasingly complete desks in selected markets. A fixed date for complete role replacement fails because roles contain changing mixtures of reproducible work, organizational decisions and capital authority. A newly introduced instrument can add settlement questions even after an agent masters the existing market.

The strongest incumbents can deploy the same software progress against their own assets. Jane Street's research description reports large data and compute resources. Our competitive inference is that automation can increase the productivity of those resources. Financing capacity, low-latency infrastructure and relationships can also remain valuable where the strategy needs them. Small teams should choose opportunities where cheaper reasoning and simpler operations matter most, rather than assume every market has become equally accessible.

The projection is bold at the task level: much of a conventional public-data, hours-to-days research workflow could become available as agent-run work this decade. The claim becomes narrower at the firm level because access, capacity and accountability shape the business. We would revise the timetable using accepted autonomous work and realized operating outcomes, rather than count job titles that still contain familiar words.