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

Loading slides…

Read the talk

How Korean Chip News Could Shape a Portfolio Decision at Coatue

Frank Long of Coatue makes his case for agents in investing by tracing one research task from collecting Korean-language semiconductor coverage to judging whether the firm should change what it owns.

From a talk by Frank Long

At a glance

Ideas worth remembering

  • Expose external sources and internal decision records through APIs. Notes on investments the firm passed on can identify the original objection that new evidence needs to address.

  • Specify what a research step must return. A source-checked quotation, a set of external figures and evidence that challenges an internal note each require a different route through the coverage.

  • Choose runners against a test the team understands. Code can retrieve a day’s articles, an LLM can extract a statement, and an investor can judge its implications for position size and timing.

A hedge fund returns three things

At Coatue, a technology investment firm active in public and private markets, an investor has to take a high-conviction bet on a theme and watch how it plays out. Frank Long thinks career decisions work much the same way. When he joined Coatue the previous year, his bet was that agents would have an even bigger impact on investing than they had on coding. To see where that impact might come from, he starts with what an investment firm actually produces.

Long writes a hedge fund as a function of two parameters, external context and internal context. It returns one tuple, (selection, size, timing). Which stock? How much? When? A correct call on a company leaves two of those questions open. The firm has to compress everything it knows about the world and its own circumstances into all three answers. Its product, in this formulation, is a time series of those tuples. Successive decisions construct the portfolio.

0:12
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

Give agents access to the evidence investors use

Long’s access principle starts with the investor’s screen. Anything a person at the firm can see in pixels, an agent inside the firm should be able to retrieve through an API. Imagine hiring the greatest investors alive and giving them computers without logins to 75 percent of the services they need. Their intelligence would do little to remedy the missing access. The same limitation applies to an agent that depends on somebody dragging in a filing or copying and pasting a news story. News, prices and filings should all be directly accessible.

A monitor displays a line chart and a table. A dotted line connects it to three API calls for Samsung news, historical prices and filings, each paired with its output fields.Open full source frame
Everything a human can see in pixels, a model can see via API

The second input is specific to the firm. What position does it already hold? What liquidity or risk requirements apply to that name? What has the firm learned from earlier investments in the sector, the stock or the management team? Two investors can read the same external report and have different reasons to act. Even within one firm, the implications of a report depend on what the portfolio already contains.

Long credits a foundation that predates his arrival. Coatue’s data science team, processes and systems have brought information together for over a decade. A separate discipline made the internal record worth retrieving. For decades, people filed notes on trades, on decisions they made and opportunities they passed on, and on what went well or poorly. A holdings list can tell an agent that the firm owns no shares. A note on a decision to pass can tell an investor which earlier judgment to revisit.

1:31
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

1:31 · section reference included

Gather evidence, form a view, then decide

Access still leaves the compression problem. Long breaks the function into three steps. Gathering takes external and internal context and filters for relevant evidence. Analysis takes that evidence and forms a view about what it means for public or private markets. The decision step takes the view and produces the object with selection, size and timing. Each output becomes the input to the next step. This gives an engineer a place to look when a sound piece of evidence fails to influence a decision.

Five rows pair investment-fund pseudocode with explanations. The function gathers information, analyzes it and returns selection, size and timing. Red annotations assign the decision and investment judgment to people.Open full source frame
Break it down until each piece can be checked

Consider the judgment already embedded in gathering. An NBA story probably has little bearing on the stocks the firm follows. A local permitting story might matter considerably if the permit affects data centers. A filter that discards local coverage can lose the relevant fact before the analysis step gets a chance to interpret it. Long narrows the problem further to see what implementing even this first step entails.

3:20
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:20 · section reference included

What reaches the model from a Korean news article?

Semiconductors and AI had become major market themes, making Korean media coverage particularly valuable to a team Long describes as entirely non-Korean-speaking. Reading that coverage gives an LLM a concrete job. The first engineering decisions come before translation. The team could collect articles through MCP, use a CLI, or license a data feed and clean it. Whichever route supplies the article, someone still has to decide how to chunk it for the model and how to represent its images. Passing an image directly and converting it to text give the model different material to interpret.

Long treats those choices like hyperparameters in a traditional machine learning system. Each can be varied and evaluated. Filtration adds another choice. Korean outlets cover basketball too, so selecting sources by language gets the team only partway to the names and themes in its portfolio. Then comes the choice of what to hand over. A translation lets an investor read the article. A summary condenses it. Structured extraction pulls particular information into a form another step can process.

The team can cross-reference that output with its own notes, existing views, conversations or meeting recordings. That changes the question from what the article says to how it bears on something the firm already believes. Updating the view is conditional. Checking the evidence may give the investor reason to keep the existing view.

Seven rows trace a research question through the gather function. Code collects articles and checks quotes. An LLM filters articles, extracts observations in English and proposes an update. A person accepts, revises or rejects it.Open full source frame
Interpreting news from Korean media outlets
4:51
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:51 · section reference included

Four kinds of runner, from a comparison to a management meeting

Long calls whoever or whatever executes a step its “runner.” His example assigns a sequence of steps to code, LLM, LLM, code and person. The broader menu has four entries. The first is deterministic code. Given a price-to-earnings ratio and a rule requiring it to be below 15, a comparison implements the rule exactly. It is fast, cheap and reliable. The second is statistics or classical machine learning. Fitting a regression line to observations lets the team estimate a relationship whose shape it would otherwise have to specify in code.

The third runner is an LLM or AI agent. Summarization is Long’s example of a task where a modern model can handle variation in language that would defeat an attempt to enumerate cases with if statements. The fourth runner is a person. For a meeting with a company’s management team, Long would send someone who can exercise judgment and navigate the encounter as it develops. He chooses each runner to maximize the absolute quality of that piece of work. The share of steps assigned to LLMs tells him little about that result.

Four columns pair a price-to-earnings rule with a black square labeled CODE, model fitting with a green diamond labeled ML, transcript summarization with a gray oval labeled LLM, and a CEO meeting with a red circle labeled PERSON.Open full source frame
You need a runner
7:15
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:15 · section reference included

Three routes through the same coverage

The research question can change how the workflow branches as well as the order of its steps. If the investor needs to know exactly what someone said, one route is to collect the articles, translate them, extract quotations and check each quotation against its source. A fluent translation alone leaves an important question open. Did the person actually say what the extraction credits to them?

A second route begins by deduplicating coverage, then filters the remaining articles and pulls external figures. Its intermediate result is a set of figures from relevant coverage. A third route reverses the starting point. The investor supplies an internal note first and asks the system to find corroborating data points. The investor could also red-team the note, looking for evidence that challenges its reasoning. Long treats control flow itself as tunable. A change in the investor’s question can justify a different first input, a different sequence and a different test of the output.

9:26
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:26 · section reference included

Evaluate a bounded task before changing its runner

Some assignments are straightforward. Pulling the last 24 hours of articles from a given source is a well-specified collection job for deterministic code. Filtering that collection might use a decision model, an LLM, a statistical method or a combination. Long expects much of the real work to involve multiple runners. A person judging portfolio impact, for example, can use a deterministic system to search for and retrieve the evidence.

Even choosing an LLM leaves decisions about a small, medium or large model, the thinking effort, the harness, which skills the harness loads and which hooks load with those skills. Long’s way into that search space is to isolate a piece of work the team can understand and evaluate. Quote extraction has a source to check. Translating a statement into portfolio impact asks for a different judgment. If an LLM does the first job well and struggles with the second, the team can keep its extraction and put the impact decision with an investor. Without a criterion at that boundary, replacing an LLM with a trained filter or an if statement offers no way to tell whether the change helped.

10:16
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:16 · section reference included

The research graph extends beyond investing

Korean news is one source, and each source may need more than three transformations just to produce an intermediate result. Long sketches the scale as at least ten tunable choices per node, each with at least twenty plausible values. For engineers, the challenge is to lay out that graph so they can understand how the fund reaches a decision. For the firm, there is also the question of how to organize the work. Someone has to decide which judgments investors should make themselves and where a combination of people and software produces a better result.

A flowchart combines Korean news, filings, call transcripts and internal notes into a proposal for review. Dashed arrows send mismatched quotes back for extraction, revisions back to the proposal, and review feedback back to internal notes.Open full source frame
Investment research workflow from collection to review

Long sees the same design problem in policy, nonprofits, drug discovery and corporate strategy. In his terms, each is building a state machine that takes information from the world, interprets it in light of what the organization knows and uses it to make coherent decisions. Investing supplies an unusually compact expression of the final decision. Much of the work lies in deciding what evidence gets there.

12:15
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:15 · section reference included

Transcript download
Full transcript

Loading transcript…