Prof. Hernan Huwyler, MBA CPA CAIO AI GRC Director · Academic Director, IE Law School
Published · Last reviewed
- Agents will absorb routine quant work first: tuning, backtest setup, stress configuration, data preparation. Judgment, objectives and accountability stay human.
- Cheap testing multiplies false discoveries; finance lacks a proof checker, so firms need a trial ledger and locked test bench.
- Junior build work shrinks, thinning future validators; train them with replication drills, and require explanations before approving any agent-built model.
On September 8, 2026, OpenAI said a group of roughly 10,000 concurrent agents had produced a solution to the Navier-Stokes problem in 88 hours. On November 20, a mathematician at Anthropic released a Lean library of about 17 million lines that the Lean and Mathlib communities certified as a correct proof of the Riemann Hypothesis. Both results came with a machine checker attached. In both cases, the human experts said that understanding the output would take far longer than producing it. Quant research and risk teams should treat these two events as a preview of their own next two years.
The direct answer is this. Agents will absorb the routine layer of quantitative work first: parameter tuning, backtest setup, stress scenario configuration, data preparation and first-draft documentation. That drives the cost of producing a research result toward zero and leaves the cost of verifying it where it was. Mathematics has a proof kernel that rejects bad arguments. Finance has nothing equivalent for a claimed edge, because a backtest that runs says nothing about whether the edge exists. So the control that matters most is a trial ledger that counts every variant an agent tried, paired with an evaluation bench the agent cannot edit.
This article is for heads of quant research, chief risk officers, model validation leads and internal auditors who will be asked to sign off on work that agents did most of. After reading it, you can write a trial-ledger policy, separate the generator from the judge, find the gap in current model risk guidance, and decide how to train validators when the junior work shrinks. Where I give numbers for a desk, they come from an illustrative scenario with stated assumptions, and I label it as such. Where I give facts about the world, they come from sources listed at the end, and I flag the ones that come from the companies selling the models.





