Frank Long of Coatue Argues AI Agents Will Reshape Investing More Than Coding

Frank Long introduced himself as a member of Coatue, which he described as a technology investment firm across public and private markets. He said that in this business, the key is being able to take high-conviction bets on big themes and watch those bets play out, and he compared a career to the same approach.

Long said that when he joined Coatue last year, the bet he wanted to make was that agents will have an even bigger impact on investing than they have had on coding. He acknowledged that this is a big statement, given the transformative impact AI has already had on coding, but said he strongly believes it and would use the talk to describe the framework the firm uses to realize that impact.

Long said that, at a high level, a hedge fund can be thought of as a function. The function takes in two parameters, external context and internal context, and produces one tuple made of three parts: selection, size, and timing. In his words, the three questions are what stock, how much, and when.

He said this framing grounds the work in a compression problem: how to take all the information in the world and produce one simple tuple. The product Coatue provides, he said, is essentially a time series of this tuple, which is used to construct a portfolio.

Long described a simple philosophy for the inputs: investors use their screens, look at information and do work, so every single thing a human can see in pixels on a screen should be visible to an agent through an API. He offered an analogy: bringing in the greatest investors alive but giving them a computer without logins to 75% of the firm's services would limit how smart the system could be if it lacked the tools to act.

External context, he said, includes news, prices, and filings, all of which should be accessible directly via API rather than through dragging, dropping, copying, or pasting. Internal context is equally important and covers the firm's own situation, such as current positions, potential liquidity and risk requirements around certain names, and past case studies or history with a given sector, stock, or management team.

Long credited KKR with leading this effort for more than a decade, describing it as both technological and cultural, with a data science team and processes and systems that bring this information in. He said the firm has filed a note on every trade and every decision it made or did not make, including what went well and what did not, and that this record has positioned it to use AI.

Long said it is not realistic to take all of these inputs, place them into a context window, and expect three numbers in return. The function, he said, has to be broken down further.

He proposed that the first step is to take in the external and internal context and gather it, filtering out what is relevant and what is not. The second step is to analyze that information and form a view on what it means for public and private markets. The third step is to turn that view into a decision object with three parameters. He noted that each step has an input and an output, and the input of each subsequent step is the output of the previous one. For an engineer, he said, this is a useful way to conceptualize the problem, but even this cannot be as simple as writing three functions.

Long said the first step, gathering, requires filtering relevant from irrelevant material, and that this involves a lot of nuance. He gave an example: if he pulled news articles about the NBA, that is probably not material to stock prices, but a story about local permitting might actually affect data centers.

To show the scope of the problem, Long said the team would look at a small piece of that gathering step using a case study on interpreting news.

Long said that news from Asia and Korean media outlets has been especially important this year, because semiconductors and AI have been a huge theme in the market. He argued that this is an area where AI is poised to have a large impact: reading Korean-language news better than a staff that does not speak Korean.

He described the collection step as one that could be done through an MCP, a CLI, or by licensing and cleaning data feeds, with many ways to do it. The step also involves many choices, such as how to chunk articles, and whether to pass images to the LLM as images or to convert them to text first. Long compared these choices to hyperparameters in a traditional machine learning system, saying each is something to tune and optimize against.

Filtration matters here too. Long said that Korean news may include many items, such as NBA coverage, that are not necessarily material to the names, themes, and stocks in the portfolio, so he wants to filter those out. He also described choices between translating, extracting, and summarizing, and whether to turn the content into a structured form. He said there are many ways to experiment with exposing information to the model.

Long said the later steps include cross-referencing information from disparate sources, which he called very powerful, and updating views. He described each line of the pipeline as having a runner, and listed four types.

Code is the most obvious runner: it is fast, cheap, and reliable. For a rule such as a price-to-earnings ratio below 15, he said deterministic code is the right choice.

Statistics and classical machine learning are the second type. He said fitting a regression can capture specific patterns and model relationships much more complex than fixed code can express.

LLMs and AI agents are the third type, suited to more complex relationships. Long said summarization is the state of the art for LLMs, and that building a set of if statements that summarizes better than a modern model would be practically impossible.

People are the fourth type. Long gave the example of a function that meets a company's management team. He said he would not want to send an LLM, an AI agent, or an if statement into that meeting, because judgment and the ability to navigate the real world make a person the best runner for that piece of the function.

Long stressed that this is an important point. He said there is no glory in automating things for the sake of automating things, and that the goal is not to make as many steps LLM calls as possible. For any given piece of the function, he said, the question is what can be done to maximize absolute quality.

Long said the function probably does not look like a linear list of steps. It probably has branching logic and many ways to fork. One approach he described is to collect information, translate it, extract quotes, and check those quotes against sources to confirm people said them. He said this might be totally wrong, and that other paths could include deduplicating and filtering, pulling external figures, finding confirmatory data points for a note, or writing a team update.

He concluded that there are a million ways to configure this, which adds yet more hyperparameters within the control flow, even for the simple problem of polling, filtering, summarizing, and aggregating news for LLMs.

Long said each box in the pipeline probably has an optimal runner. For collection, that might be deterministic code, such as polling the last 24 hours from a given source. Filtering or decision steps might call for models, LLMs, statistical methods, or a combination.

He said this understates the complexity, because most things are multi-runner. Many steps combine deterministic code, an LLM, a statistical method, or a human working with an LLM. Within the LLM category there are further choices: which model size to use, how much thinking effort to apply, which harness to use, and which skills and hooks to load with it. Long said the search space therefore contains an enormous number of hyperparameters, even for the single tiny task of filtering news.

Long said the way to break down these problems is to find what can be compartmentalized in a way that can be understood and evaluated, because that is what the team hill-climbs against. Without that, he said, the work is just tuning. For each piece, the question becomes whether it should be a machine learning system, a small model the team trains, an if statement, or no model at all.

He said LLMs are really bad at translating news into portfolio impact, so that step should probably be a human, supported by a deterministic system that helps search and pull the information. He concluded that without a North Star to hill-climb against, you are just fidgeting around.

Long said the example he presented is just one tiny thing the firm does. Every data source pulled in likely needs more than three transformations to reach an intermediate output, so the full graph for a fund like Coatue is massively complex.

He said each of those nodes has at least 10 hyperparameters, each with at least 20 valid values, so the search space is basically infinite. He described laying this out so it can be comprehended, both technically and culturally and organizationally, as the great challenge of applying AI in investing.

Long said that what the talk is really about is creating a state machine that can take information from the world and understand it well enough to make coherent decisions. He said such use cases reach far beyond investment firms, naming policy, nonprofits, drug discovery, and corporate strategy.

He said one of the most exciting parts of being at events like this is talking with people in different spaces who have different problems that converge on the same problem space, and he offered to continue the conversation in the hallway afterward.