# AI Agents Will Upend Investing Even More Than Coding

Frank Long, Coatue | AI Engineer NYC 2026 | 14:05

Source: https://www.youtube.com/watch?v=AYFcmnQrZCg
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/ai-agents-will-upend-investing-even-more-than-coding
Published: 2026-11-06
Tags: agents, evals, finance, workflows

## TL;DR
- Frank Long models a hedge fund as a function that turns external and internal context into a stock selection, position size, and timing decision.
- Agents need API access to everything people can see on their screens, including news, prices, filings, positions, risk constraints, and the firm's past decisions.
- Each step should use whatever runner produces the highest quality, whether that is code, statistics, an LLM, a person, or a combination of them.

## Summary
Frank Long argues that agents could affect investing more than coding because investing compresses a huge amount of information into a small decision: what to buy, how much to buy, and when. He describes a hedge fund as a function that takes external context, such as news and prices, together with internal context, such as current positions, liquidity, risk requirements, and past research. The function must be broken into gathering, analysis, and decision steps. A Korean semiconductor-news example shows how many choices appear even in gathering information: data sources, chunking, images, translation, filtering, summarization, cross-referencing, and updates. Long assigns each part a runner, which may be deterministic code, statistics, an LLM, a person, or a combination. He is clear that automation is not the goal. Teams need a quality measure to tune against, because the resulting graph contains a very large search space of models, prompts, control flow, and human involvement.

## Key ideas
### A hedge fund compresses broad context into three decisions
[00:47](https://www.youtube.com/watch?v=AYFcmnQrZCg&t=47s)
Long describes a hedge fund as a function with two inputs, external context and internal context, and one output tuple: selection, size, and timing. The output answers what stock to buy, how much to buy, and when. A portfolio is then a time series of these decisions. This framing gives engineers a concrete target for the information-processing system: take information from the world and the firm's own situation, then reduce it to an actionable investment decision.

### Agents need the same computer access as the people they replace
[01:31](https://www.youtube.com/watch?v=AYFcmnQrZCg&t=91s)
Long's rule is that anything a human can see in pixels should be available to an agent through an API. He compares an agent without access to company services to the smartest investors being given computers without logins to 75% of the systems they need. External inputs include news, prices, and filings. Internal inputs include current positions, liquidity, risk requirements, and the firm's history with a stock, sector, or management team. Coatue's long record of notes about trades and decisions gives its systems material to work with.

### The investment function has gathering, analysis, and decision stages
[03:28](https://www.youtube.com/watch?v=AYFcmnQrZCg&t=208s)
Long breaks the large function into smaller functions. Gathering means collecting information and filtering out material that does not matter, such as unrelated news, while finding items such as local permitting that could affect data centers. Analysis turns the gathered information into a view about public and private markets. The final stage turns that view into a decision object containing selection, size, and timing. Every stage has an input and an output, with each output feeding the next function.

### A Korean-news workflow exposes many design choices
[04:59](https://www.youtube.com/watch?v=AYFcmnQrZCg&t=299s)
Long uses Korean-language semiconductor news as a small case study. Even collecting the articles has several options: an MCP, a CLI, licensed data feeds, and different cleaning methods. The system must decide how to chunk articles and whether to pass images directly or turn them into text. It must filter irrelevant material, translate, summarize or extract structured fields, and cross-reference articles with internal notes, views, and meeting recordings. Long treats these choices like hyperparameters in a traditional machine-learning system.

### Different parts of the workflow need different runners
[07:24](https://www.youtube.com/watch?v=AYFcmnQrZCg&t=444s)
Long names four types of runners. Deterministic code is fast, cheap, and reliable for rules such as checking whether a price-to-earnings ratio is below 15. Statistics and classical machine learning can model more complex relationships. LLMs are useful for tasks such as summarization. People remain the right runner for work requiring judgment and interaction with the real world, such as meeting a company's management team. A workflow can combine several runners within one function.

### Quality matters more than replacing people with models
[09:08](https://www.youtube.com/watch?v=AYFcmnQrZCg&t=548s)
Long rejects automation as an end in itself. The goal is to maximize absolute quality for each part of the investment function, rather than turn as many steps as possible into LLM calls. The workflow may include branching paths. One approach might translate a document, extract quotes, and check them against sources. Another might deduplicate, filter, and retrieve external figures. The right sequence depends on the task being optimized.

### The real system is a large search space that needs an evaluation target
[09:26](https://www.youtube.com/watch?v=AYFcmnQrZCg&t=566s)
Each workflow has many tunable choices. A system may combine code, statistical methods, an LLM, and a person. Within the LLM part, teams can choose model size, thinking effort, harness, skills, and hooks. Long says teams need to compartmentalize the problem so they can understand and evaluate it. Without a clear quality target, changing a model, an if statement, or a human handoff becomes fidgeting rather than measured improvement.

### Investing is one example of a broader decision-making state machine
[12:38](https://www.youtube.com/watch?v=AYFcmnQrZCg&t=758s)
Long says the full system is a massively complex graph rather than a simple sequence. He describes the broader problem as creating a state machine that takes information from the world and understands it well enough to make coherent decisions. He sees related applications in policy, nonprofits, drug discovery, and corporate strategy. The technical and organizational challenge is making such a graph understandable and usable inside an institution.

## Notable quotes
- "Every single thing that a human can see in pixels on the screen, an agent in your building should be able to see via API." (01:31)
- "There is no glory in automating things for the sake of automating things." (09:08)
- "The goal is, for any given piece of this function, what can I do to maximize absolute quality?" (09:08)
- "And if you don't have a North Star to hill climb against, you're just kind of fidgeting around." (11:43)

## Tools & references mentioned
- Coatue
- MCP
- Unitree robots

## Who should watch
- You are building agents for research or investment work and need a way to decide which steps should be handled by code, models, or people.
- Your agent has access to documents but lacks the firm's internal positions, risk constraints, past decisions, or other context needed to act well.
- You are tuning a multi-step workflow without a clear quality measure and want a concrete example of how the search space grows.

## Related talks

- [AI Agents Cut 3-Day Portfolio Builds to Minutes](https://aietalks.com/talks/ai-agents-cut-3-day-portfolio-builds-to-minutes) (Michael Manapat, Rowspace & Jack McDonald, Apollo / Redding Ridge, 18:57)
- [Agentic Engineering: Working With AI, Not Just Using It](https://aietalks.com/talks/agentic-engineering-working-with-ai-not-just-using-it) (Brendan O'Leary, Kilo Code, 27:03)
- [Building Applications with AI Agents](https://aietalks.com/talks/building-applications-with-ai-agents) (Michael Albada, Microsoft, 15:50)
- [Scaling Agents for Gen AI Products](https://aietalks.com/talks/scaling-agents-for-gen-ai-products) (Anju Kambadur, Bloomberg, 19:38)
- [How agents will unlock the $500B promise of AI](https://aietalks.com/talks/how-agents-will-unlock-the-500b-promise-of-ai) (Donald Hruska, Retool, 16:22)
