MEDALLION FUND, CODING AGENTS, AND THE NEW ECONOMY
Quantitative finance as a precedent for AI-native work
In 1988, Jim Simons launched the Medallion Fund, which would go on to become the most successful investment vehicle in history, averaging annual returns of 66% before fees over three decades. The deeper revolution, however, was not financial but organizational. What Simons proved was that the highest-leverage human contribution in a complex domain is not executing the trades but specifying what the machine should optimize for. The trader who reads charts and places orders was replaced by the researcher who designs the model, defines the objective function, and sets the risk constraints, leaving the machine to handle everything downstream.
Three decades later, the same structure is appearing in software. A developer sits down with a coding agent, describes what they want built, specifies the constraints and acceptance criteria, and watches the machine write code, run tests, diagnose failures, and iterate toward a solution. What they are doing is not programming in any traditional sense but rather specifying an optimization target and watching a machine explore the solution space, much as a quantitative researcher specifies a trading strategy and watches the model execute it.
The Researcher and the Trader
The organizational insight of quantitative finance is frequently misunderstood as a shift from intuition to mathematics, when the real transformation concerns the structural relationship between human judgment and machine execution. A quantitative researcher does not trade directly. Instead, they design a model, specifying which signals to attend to, what objective function to optimize, what risk constraints to enforce, and what evaluation criteria will determine whether the model is performing as intended. The machine then explores the solution space, executing thousands of trades across hundreds of instruments at speeds and scales no human could match, while the researcher monitors performance and revises the specification when results diverge from expectations.
The quantitative researcher's edge lies in the ability to formalize a view about market structure into a specification precise enough for a machine to act on, while leaving the machine free to discover the specific actions that satisfy the specification. The human contributes judgment and domain knowledge about what matters, while the machine contributes speed, scale, and tireless exploration of possibilities.
Coding agents instantiate this same relationship, only now applied to software rather than markets. The developer specifies intent, constraints, and acceptance criteria, then watches the agent explore the solution space by writing code, running tests, diagnosing failures, and iterating toward something that works.
Tight Loops and Small Teams
The most effective AI-native teams already resemble trading desks in their structure and operating rhythms. They are small, hyper-specialized groups where each member formulates problems for machines to solve rather than solving them manually. The hallmarks of this operating model are recognizable in both settings. Feedback loops are tight, with the consequences of a specification visible in minutes rather than weeks. Experimentation is cheap and reversible, organized around branches that can be discarded without cost if they fail to produce results. The culture prizes the quality of the question over the speed of the answer, understanding that a well-formulated problem is already halfway to its solution.
The Boundary
This topology holds wherever work can be decomposed into specification and execution connected by measurable feedback. A data analyst defines the question and lets the machine run the queries. A designer specifies a visual direction and evaluates generated alternatives against it. A content producer sets engagement targets and lets automated systems optimize distribution toward those targets. Any domain where someone can build a rapid evaluation loop, where the machine can test whether a proposed solution meets the specification quickly and cheaply, becomes a domain where this structure will eventually prevail.