AI Agents Will Upend Investing Even More Than Coding - Frank Long, Coatue

A Korean-language report about a semiconductor company lands in front of an investment team that doesn't speak Korean. An LLM can help the team read it. The harder question is whether anything in the report should change the firm's position. That depends on what the firm already owns, what it previously believed about the company and how much risk it can take.

Frank Long, who works at Coatue, a technology investment firm active in public and private markets, describes a hedge fund as a function with two inputs and three outputs. The inputs are external information and the firm's internal context. The function returns a selection, a size and a time: which stock, how much and when. A succession of those decisions constructs a portfolio. In that framing, a system that correctly interprets the news has solved one part of a much larger compression problem.

Long's bet is that agents will have an even bigger impact on investing than they have had on coding. He presents that as a thesis. His case rests on how much work sits between receiving a piece of information and making a decision, and how many of those intermediate tasks an engineer can isolate and improve.

Long's design principle is that information an investor can see on a screen should be available to an agent through an API. Prices, news and filings are obvious candidates. If an investor has to find a filing and paste it into a chat window, the investor is still doing the gathering. The same access problem applies inside the firm. A view on a company has different implications depending on the firm's existing position, its liquidity requirements and its experience with that management team.

Building an API solves only the retrieval half of that problem. Long says Coatue has spent decades recording notes about trades, decisions it made, decisions it passed on and what went well or poorly. A list of holdings tells an agent what the firm bought. A note on an investment it declined can tell the agent what objection mattered at the time. That gives the research process something specific to check when a new report arrives.

Long breaks the larger function into gathering evidence, analyzing it to form a view and turning that view into a decision. Even gathering involves judgment. An NBA story is unlikely to matter to the portfolio. A local permitting story might matter a great deal if the permit affects a data center. A filter has to recognize that connection before the analysis step ever sees the story.

In Long's Korean news example, the first choices arise before translation. The team could collect articles through a command-line tool, an MCP interface or a licensed feed that needs cleaning. It must decide where to split an article and whether to send its images to the model directly or convert them to text. Those choices determine what evidence survives collection. A downstream model can work only with what the collector passes along.

Translation, summarization and structured extraction also produce different handoffs. If the question is what an executive said, Long sketches a route that translates the article, extracts the relevant quotation and checks it against the source. To decide whether the report changes the firm's view, the team can then compare it with internal notes, earlier conversations or meeting recordings. Updating the view is a distinct, conditional step. Sometimes the work ends with the comparison.

Long calls the executor of a step its “runner.” His simplest example is a price-to-earnings ratio below 15. Once the ratio is available, a comparison in code is fast, cheap and reliable. A relationship that needs to be estimated from observations may call for a statistical method such as regression. Summarizing a varied collection of articles is a better fit for an LLM than attempting to encode the task as a series of conditions.

The fourth runner is a person. Long would send one to meet a company's management team because judgment and the ability to navigate the encounter matter to the result. Most steps, in his account, can combine runners. Code can retrieve the last 24 hours of coverage from a specified source. A person can use that retrieval system to find evidence and judge its portfolio impact. Long's criterion is the quality of the result, even when the best design keeps the decision with an investor.

The quote-checking route is only one way through the Korean coverage. Long describes another that removes duplicates, filters the articles and pulls external figures. A third starts with an internal note and looks for corroborating evidence. An investor trying to challenge that note would give the search a different purpose. All of these might appear on a project plan as news analysis, yet each asks the system to return something different.

Those differences give an engineer a way to judge a change. For the quotation task, the source provides a concrete check: did the speaker actually say what the extraction attributes to them? That check says little about an investor's ability to translate the statement into a position. Long argues for drawing boundaries around pieces of work the team can understand and evaluate. Otherwise, switching from an LLM to a trained filter, or moving a judgment call back to a person, is just another configuration change.

Model size and reasoning effort add choices inside a single step, as do the harness, skills and hooks loaded around the model. Multiply those choices across the transformations for every data source and the research process becomes a large, branching graph. Long sees an organizational challenge here alongside the technical one. Investors need to understand what a step contributes well enough to judge whether a change improved the evidence they use. A source-checked quotation gives them something concrete to inspect before anyone asks the system to recommend a trade.