Frank Long on Building AI Investment Workflows Around Quality

I work at Coatue, a technology investment firm active in public and private markets. My bet is that AI agents will eventually have an even bigger impact on investing than on coding. That is a strong claim; what I want to explain is the framework we use to try to realize that impact.

Think of a hedge fund as a function. Its inputs are external and internal context; its output is a decision with three components: what to select, how much to hold and when to act. A portfolio is a time series of those decisions. The challenge is compressing an enormous amount of information into that small decision.

My first principle is that information a person can see on a screen should also be accessible to an agent through an API. Even an exceptionally capable investor would be hamstrung if they joined a firm and lacked access to most of its tools. Intelligence alone is not enough when the system cannot reach the information it needs.

External context includes news, prices and filings, ideally available directly rather than through manual copying. Internal context matters just as much: current positions, liquidity and risk constraints, past case studies, and the firm's history with a sector, company or management team.

The foundation is cultural as well as technical. Data systems make information available, but a rigorous process for recording what was considered, what was done, what was not done and what happened afterward makes accumulated institutional knowledge usable. Decades of notes create a potential resource for AI systems.

I cannot simply stuff all of that context into a model and expect it to produce three magic numbers. I break the problem into gathering relevant material, analyzing it to form a view, and turning that view into a decision. Each stage has inputs and outputs, with the output feeding the next stage.

Even gathering is nuanced. A sports story may have little to do with a position, while a local permitting story could matter to a data-center theme. Once relevant material is found, the team still has to interpret its implications for public and private markets and decide how it affects the portfolio.

To see the complexity, consider one small task: interpreting Korean-language news about semiconductors. The sector has been important to AI investing, and language models may help a team that does not speak Korean read the relevant coverage. The usefulness of this approach still has to be tested.

Collection alone offers many choices: license and clean feeds, access them through an MCP server or command-line tool, decide how to split articles, and decide whether images should be passed to the model as images or converted into text. I think of these design choices as parameters to tune.

Next comes filtering: much of the news will not matter to our positions and themes. Then we choose whether to translate, summarize or extract a structured record. The best way to expose information to a model may also change as the models themselves change.

We also need to cross-reference the outside information with our own notes, views and conversations. Pulling evidence from different sources and checking it against other evidence can be powerful; it can also lead us to update an existing view rather than merely add another summary.

Each step can have a different runner. Deterministic code is fast, inexpensive and reliable for an explicit rule. Statistical methods and classical machine learning can model more complicated relationships. Language models and agents provide another option for tasks such as summarization.

People are runners too. A meeting with a management team may call for human judgment and the ability to navigate the real world. There is no value in automating a step merely to increase the number of automated steps: the objective is the best quality for the function as a whole.

The workflow is not necessarily a straight line. One design might translate, extract quotations and verify them against sources; another might deduplicate, filter and pull additional figures. A different starting point might be an internal note that needs supporting or contradictory evidence.

Different stages may combine runners instead of selecting just one: code with a model, statistical filtering with an LLM, or a human using deterministic search. Even within the LLM option there are choices of model size, reasoning effort, harness, skills and hooks.

The search space becomes huge even for news filtering. To improve it, we need to isolate pieces we can understand and evaluate, and define the outcome we are optimizing for. Without that north star, swapping an LLM for a classifier or a rule is just tinkering.

The right answer can differ by stage. A model may be weak at translating an item into portfolio impact, so a human might make that judgment with deterministic tools helping them find evidence. The point is to choose the combination that works, not to privilege any one technique.

The real graph across all the firm's data sources is much more complex than the news example. Each node has multiple design choices, and the organization has to understand the system technically and absorb it culturally. That is the practical challenge of applying AI to investing.

The same pattern extends beyond investing: a system takes information from the world and tries to turn it into coherent decisions. Policy, nonprofits, drug discovery and corporate strategy face versions of that problem too, so methods and lessons can travel across fields.