Coatue’s Frank Long says AI agents will reshape investing more than coding, and warns against automating everything

Add preferred source
Frank Long, who joined Coatue in 2025 to build AI into the firm’s investment process, argues that AI agents will have a bigger impact on investing than they did on software engineering — a bold claim from someone who has watched coding agents upend their own industry, according to remarks made on the AI Engineer podcast. His case rests on a simple model: a hedge fund is a machine that turns information about the world into three numbers — which stock, how much, and when — and most of the work is compressing an overwhelming flood of data into those three outputs. Long lays out a three-stage pipeline (gather, analyze, decide) and then shows that even the smallest sub-step, like reading Korean-language semiconductor news, hides hundreds of tunable choices that multiply into what he calls an effectively infinite search space. The solution, he says, is not to hand everything to a large language model but to match each piece of the job to one of four “runners” — deterministic code, classical statistics, LLMs, or humans — and to hill-climb against a clear quality target. The wider implication for investors: the competitive moat in AI-driven investing may be a decade of internally filed notes on every trade and decision, an asset that cannot be bought or scraped, while the industry’s $730 billion-plus AI buildout continues to feed the very data pipelines those agents will consume.

Key Elements

Frank Long’s claim sounds like hype until you consider where he sits. He joined Coatue, the technology-focused investment firm, in 2025 to build AI into the firm’s investment process. And on the AI Engineer podcast, he made a bet that is easy to dismiss and hard to ignore: “I think that agents will have an even bigger impact in investing than they did at coding. I know that’s a big statement.”

It should be. Coding was the first place where AI agents went from a demo to something companies actually pay for. If Long is right, the next wave will not produce faster autocomplete for software engineers — it will produce something closer to an autonomous research analyst that reads the world, forms a view, and tells a portfolio manager what to do.

The reason to take him seriously is not the prediction itself but the framework behind it. Long spends most of the interview dismantling the naive version of the idea — that you simply point a large language model at a Bloomberg terminal and let it trade — and replacing it with something more mundane and more useful: a blueprint for rebuilding a fund’s entire information workflow around AI, one small step at a time.

A hedge fund is a machine that squeezes the world into three numbers

Long’s organizing insight is almost embarrassingly simple. A fund takes two inputs — external context, meaning everything happening in the world, and internal context, meaning the fund’s own positions, risk limits, and history — and produces one output: selection, size, timing. Which stock, how much, when.

Everything else is compression. The job is to turn a firehose of information into a time series of those three numbers that builds a portfolio. He splits that job into three stages, each feeding the next:

The input rule is blunt, and it is where most AI projects quietly fail: “Every single thing that a human can see in pixels on the screen, an agent in your building should be able to see then handing them a computer with no logins to 75% of the company’s services. The intelligence is nearly irrelevant without the tools to act on it

That means external data — news, prices, filings — has to arrive by API, not by drag-and-drop or copy-paste. Internal context matters just as much: what the fund already owns, how much liquidity it needs around specific names, and the history of past bets on a sector, a stock, or a management team.

Long credits the firm, not himself, for the foundation. Coatue’s data science team has been building these systems for more than a decade, he says, but the harder asset is cultural. For decades the firm filed a note on every trade, every decision made or not made, and what went well and what didn’t. That archive, he argues, is what actually positions a fund to expose itself to AI — and it is the part no competitor can buy or scrape.

Even “read the news” hides an infinite number of dials

Here is where Long’s talk gets genuinely useful for anyone who thinks AI is just a model choice. He takes a trivial-sounding task — understanding the news — and shows it exploding into dozens of tunable decisions he calls hyperparameters.

He chooses Korean-language media as his case study for a specific reason: Asian and Korean outlets have been important to the semiconductor and AI themes in 2026, and AI reading Korean better than a non-Korean-speaking investment staff is a plausible edge. Watch what a seemingly simple pipeline actually requires:

Each of these is a dial. Collection alone offers MCP, a command-line tool, or a licensed and cleaned data feed; chunk articles one way or another before passing them to a model; decide whether to pull images as images or convert them to text. Filtration means stripping out noise — his example is Korean basketball fandom content with zero bearing on portfolio names. Translation, extraction, and summarization branch again, including whether to force the output into a structured form.

Then there is cross-referencing, which he calls “super powerful”: checking incoming articles against internal notes, prior views, conversations, and meeting recordings, and potentially revising the firm’s view.

And control flow adds another layer entirely, because the pipeline is not linear — it branches. One configuration might collect, translate, extract quotes, and verify them against sources. Another might dedupe, filter, then pull external figures. Another might start from an existing note and hunt for confirmatory data, or run a red-team pass to try to break the thesis.

Within the LLM bucket alone there are more dials: model size (big, small, medium), thinking effort (extra-high, high, medium), which harness to use, which skills load with the harness, which hooks load with the skills. His conclusion is that the search space is “basically infinite,” and that the real graph of a fund like Coatue is far more complex than his three-stage sketch. He jokes that the simple version is his screensaver. Each node, he says, carries at least ten hyperparameters with at least twenty valid values.

Four kinds of “runner,” and no prize for automating for its own sake

Long’s most quotable line is a warning to the automation-obsessed: “There is no glory in automating things for the sake of automating things. The goal is not to take as many things and make them LLM as possible for us. The goal is, for any given piece of this function, what can I do to maximize absolute quality?”

That principle produces a four-way taxonomy. Every step in the pipeline should be assigned to whichever runner does it best:

That last row is not a joke. He name-checks Unitree’s new robots while explaining why he still wouldn’t send one into a management meeting. Judgment, reading a room, and probing a CEO’s evasions remain human work.

The practical takeaway is that most real steps are not one runner but a combination — deterministic code plus an LLM plus a statistical method, or a human working alongside a model. Getting this wrong in either direction is costly. LLM-maximalism wastes money and reliability on tasks that simple rules handle better. Automation-minimalism leaves the hardest analytical work on the table.

Without a North Star, you’re just fidgeting

If the configuration space is effectively infinite, how do you ever finish tuning it? Long’s answer is discipline: break the problem into pieces small enough to understand and evaluate, and define a quality target for each one. That target is what he calls the North Star, and without one “you’re just fidgeting around.”

The tuning loop is empirical, not theoretical. You discover through measurement that a step should be classical ML instead of an LLM, or a small trained model, or not a model at all but an if-statement. You might find that LLMs are bad at translating news into portfolio impact, so a human should own that step while a deterministic system does the searching and pulling.

He is careful to note that the news pipeline is one tiny piece of the machine. Every databefore it reaches a usable intermediate output, and the full graph is massively complex. Laying that graph out so it can be understood technically and absorbed culturally and organizationally is, in his words, the great challenge of applying AI in investing

Why the whole economy faces the same state-machine problem

Long generalizes the fund problem to any organization that must take information from the world, understand it well enough, and make coherent decisions — a state machine for decision-making. He names policy, nonprofits, drug discovery, and corporate strategy as domains with the same shape, and says the most exciting part of being in forums like this one is meeting people from other fields whose problems converge on the same space.

That generalization matters for investors because it reframes where the value will accrue. The companies that win in AI may not be the ones with the biggest models. They may be the ones with the cleanest internal data, the most rigorous habit of recording decisions, and the patience to decompose their workflows into evaluable pieces.

Meanwhile, the industry is spending at a pace that makes the stakes enormous. Global technology AI spending is set to exceed $730 billion this year, and U.S. hyperscalers and related companies issued about $225 billion of bonds in the first nine months of 2026, up nearly 974% from a year earlier That buildout produces the exact data deluge Long’s pipelines are designed to consume — and the capital markets are already asking whether the returns will justify it

The unresolved tension in Long’s argument is the one he states rather than solves: the search space is effectively infinite, the right decomposition is not obvious in advance, and the only defense is disciplined compartmentalization plus a measurable target. Whether agents deliver more in investing than in coding remains his bet, not a demonstrated result. The Korean-news pipeline is the smallest concrete instance he offers — a single box in a graph too complex to draw.

For investors watching this space, the most important thing to track may be less glamorous than model benchmarks. It is whether firms treat their internal archive — decades of notes on trades and decisions — as the real moat. That input cannot be bought or scraped, and in Long’s framework it is what separates a fund that can teach an agent to think like its best investors from one that can only teach it to read the news.

Full content available at:AI Agents Will Upend Investing Even More Than Coding – Frank Long, Coatue

• State Street CEO Ron O’Hanley: AI Investment Backed by Real Orders, Hard to Call It a Bubble at This Stage
• Howard Marks Flags a Credit Trap Inside the AI Boom: Lenders May Be Paid Too Little for Too Much Risk
• Nikkei Plunges 2,600 Points as US Midterm ‘Blue Wave’ Stokes AI Spending Fears
• Blackstone, Goldman Say AI Boom Needs Mega-Fundraisings as Spending Tops $730 Billion