Agentic Trading

In agentic trading, a computer program buys and sells stocks or other securities largely on its own. What’s special here isn’t that a machine is trading — that’s been happening for decades. What’s new is that the program doesn’t simply work through a fixed list of if-then rules. Instead, it’s given a goal, such as: “Unwind this position over the course of the day without pushing the price down.” How it achieves this goal is something it decides on its own along the way. This is made possible by AI systems that can analyze text and numbers and derive a sequence of actions from them.

Why banks are paying attention

A trader at a large bank spends a considerable part of their day not making decisions, but doing preparatory work. They read news, check quarterly figures, compare prices, and pull data from various systems. This exact chain of tasks can be handed off to a program that works around the clock and never gets tired. So the promised benefit lies first and foremost in speed and reduced staffing costs.

There’s a second factor as well: volume. A human can perhaps keep an eye on twenty markets, while a program can monitor thousands simultaneously. That’s why large asset managers like BlackRock or trading firms like Citadel are investing heavily in such systems. Trading software providers, too, are now openly marketing AI agents to retail investors.

At the same time, this is causing concern among regulators. If many programs are built similarly and react the same way to the same piece of news, they amplify market movements. A price drop can snowball within minutes as a result. The European Securities and Markets Authority (ESMA) and the Bundesbank have repeatedly warned about exactly this.

From goal to executed order

At the core is usually a language model — an AI system trained on vast amounts of text that uses this training to formulate responses. On its own, however, this model can’t buy anything. That’s why it’s given tools to work with: access to price data, a news search function, a calculator, and an interface to the broker’s trading system. The model decides which tool to use and when.

The process runs in loops. The system assesses the situation, chooses an action, carries it out, and reviews the result. Then the cycle starts over, incorporating the new information. Experts call this pattern “plan, act, observe.” Because each cycle builds on the previous one, the system can also handle multi-step tasks for which a single instruction wouldn’t be enough.

Limits are almost always hard-coded in as well, known as guardrails. Typical examples include a maximum amount per order, a loss threshold at which everything is halted, and a list of approved securities. For larger sums, a human must additionally give approval before the order goes out. This isn’t a minor detail — it’s the decisive difference between an experiment and a product that’s permissible under regulatory rules.

Between broker app and headline

In Germany, fully automated agentic trading is barely available to retail customers so far. What broker apps offer is mostly a preliminary stage: an assistant that comments on portfolios, summarizes news, or suggests orders. The final click is still made by the user. Robo-advisors, in turn, do invest automatically, but according to fixed rules — which is why they don’t count as agentic trading.

In the news, the term mainly comes up in two contexts. First, when financial firms announce new AI products and their share prices react accordingly. Second, when regulators discuss rules for such systems, for example as part of the EU AI Act.

One common misconception is worth clearing up in closing: agentic trading is not the same as high-frequency trading. That’s about millionths of a second and very simple rules, whereas this is about complex decisions made over minutes or hours. And anyone entrusting their own money to such systems should know that no one can fully trace or understand their decisions.