Irene Aldridge wrote the book on high-frequency trading—literally, the Wiley textbook that's become the industry standard.

Now she says that AI agents are the next step in trading evolution.

Her new paper, "Agentic Artificial Intelligence in Finance: A Comprehensive Survey," argues that we've crossed a threshold from AI that predicts to AI that acts. She published the new research with her students from Cornell University, where she's been a visiting professor for the last nine years. 

AI agents are autonomous systems with objectives, memory, and the capacity to adapt—agents that decide what to trade, execute, and learn from the outcome. This differs from “traditional” AI which has typically relied on human-written code.

The practical result, she argues, is systems that mirror how trading floors actually work: one agent maximizing profit, another minimizing risk, a third handling compliance. "Each agent has a different objective, but together they make the whole stronger because they control each other's behavior."

In a recent conversation, we dive into what this means for the future of markets:

This interview has been edited for clarity and length.

Matt Robinson: Your background's with HFT using deterministic rules for trading. How is generative AI different?

Irene Aldridge: Up to about 10 years ago, machine learning was much about trial and error: "Let's try this route. If it doesn't work, let's switch our machine to this route." Computer scientists were mostly thinking about speeding up machine learning by getting faster hardware.

Recently, optimization got involved. Optimization is a different department from Computer Science ... It is Operations Research. Now, Operations Research works on dramatically speeding up algorithms and finding high-precision solutions -- something I have always been very passionate about.

Matt Robinson: How does that change the approach?

Irene Aldridge: Instead of just having computers run different paths to try to find optimal solutions, now you set an objective: "This is what I want to accomplish." And let's find the optimal path to fulfilling that objective. Let's optimize it across all the different dimensions. And this is what really gives birth to these agents.

These agents have an objective in mind. And this can be a multi-period objective. So we no longer build static, globally parameterized solutions. It's no longer just a simple equation, simple function. We're looking for a dynamic system that works its own thing, by optimizing its own experience through time.

These optimization techniques have been around since World War II. Now we're giving this same approach to machine learning systems with specific goals—they recalibrate themselves as they go.

Matt Robinson: Is it like gradient descent?

Irene Aldridge: Gradient descent is one of the techniques available to create these dynamic systems. Gradient descent allows you to move toward your target minimum or maximum. In agents, you do this continuously along with other techniques to find your optimal goal.

Irene Aldridge: Now, instead of having multiple [human] traders, you have multiple automated agents performing trader functions. Some agents always seek to maximize profitability. But that's not enough. You need agents to control the risk as well. So you have compliance agents that work alongside the trader agents to contain risk, just like on a traditional trading floor!

The idea of cooperating or even adversarial agents is not unique to trading. Networks of adversarial agents, specifically, "Generative-Adversarial Networks (GANs)" are fundamental blocks of the modern Artificial Intelligence models like ChatGPT. We are expanding and adapting AI models to closely imitate traditional financial structures.

Right now, AI trading agents have an advantage: they're mostly trading against humans, and humans are predictable. You can assume that humans who are trading are profit-maximizing—they're always trading only to maximize profits.

But that's about to change. As more agents come online, they'll increasingly be trading against each other—and agents don't behave like humans.

Part of the problem is that agents are designed to occasionally act randomly. To avoid getting stuck in patterns, they'll "go off on a random tangent just to do something completely wild, just to try something new." Other agents observe those random decisions and take them as input—which generates noise throughout the system. Now you have to distinguish what's random and what's not random. And it's really hard.

And what if someone configured their agent to destabilize something? How do you actually identify them and separate them from the rest of the agents that are profit-maximizing?

That's a real challenge. I think that should be regulators' biggest concern.

Matt Robinson: Have you ever seen an experiment where they said, "Regenerate this image of the Rock," and they did it like a hundred times? By the end of it, it was complete gibberish.

Irene Aldridge: You have a random number generator that creates arbitrary stuff. And then you have these discriminators who critique whatever decisions were made by the random number generator and adapt them to make them better. That's really the essence of large language models or agentic AI. You have some innovation always filtering through the system, so it's not static—it allows it to respond to new things.

Agents continuously intake all environmental feedback and any data they can find. That allows them to be very adaptive. Traditional systematic traders had rules and checked if the environment matched those rules. Agents try to classify the world by breaking it down into certain states—using what's called a Markov Decision Process—creating a map of where you've been and where you are now, and determining what action is best for that particular state.

Matt Robinson: I've been hearing more about the demand for compute and how there's an arms race for compute. Because speed is only—you can only go so fast. Everyone is trying to get a ton of compute, how are they using it?

Irene Aldridge: I subscribe to a different point of view. I've always been bootstrapped, doing everything on a shoestring—I don't like to spend a lot of money. So there are two ways you can approach it: you either increase your computing infrastructure, build a bigger array of processing, or you use math to optimize your process so you don't have to invest a huge amount into computing.

I'm presenting a paper at a simulation conference where we show that even traditional Monte Carlo methods—the kind everyone uses in risk management—can be improved by a factor of 10,000 just through mathematical optimization. This is math that's been sitting in academic departments as intellectual exercises. It hasn't made its way into computer science or finance yet.

My preference is to make things fast with math rather than brute force. Instead of "let's buy more chips," it's "let's be smarter about the problem."

Matt Robinson: I covered finance and financial regulation... it seems like with AI there's not much of that right now. How do you see that playing out?

Irene Aldridge: Regulators need AI to combat AI. They can't have people managing this. AI is sophisticated enough that they can have their own agents that will police that. Regulators don't have a leg to stand on right now—they're not prepared at all. But it's not that hard to build these things. They have to be purpose-built systems.

You have compliance systems that are purpose-built, and you'd have regulator AI that's purpose-built. It feeds into reg-tech, but more advanced. We just need time and resources to make this happen.

Matt Robinson: What's your sense of actual implementation? My sense is that nothing's really live right now.

Irene Aldridge: There are a few very successful organizations trading with AI agents. I think right now these early AI agent adopters are very successful because there aren't as many agents yet in the markets. These agents are able to learn directly from the human environment—human actions are well thought out. People that are trading there, they generally trade because they researched their trades. So they're not making random choices.

However, when many agents come online... that's when it's bound to be wildly unpredictable, because agents will have to learn from random choices of other agents and distinguish randomly generated choices from human actions -- all very complex tasks. Regulators need to start building and adopting agentic compliance now. Soon, it may be too late.

Matt Robinson: You have a trading system running?

Irene Aldridge: Yes, I do have a trading system running. I am on track to 30% net (after costs) return for this year with almost 0 volatility (beating the S&P 500 by 11% year-to-date just trading cash equities and ETFs). It is a great diversifier to other alternative strategies. Looking to expand in the 2026!