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

## Executive summary

Frank Long posits that AI agents will revolutionize investing even more profoundly than AI impacted coding. He models a hedge fund's decision-making process as a function that takes external and internal context to output a selection, size, and timing for investments. The core technical insight is that this complex function must be broken down into sequential steps (Gather, Analyze, Decide), each requiring a specialized 'runner' (Code, Statistics, LLM, or Human). The goal is not maximizing automation, but maximizing absolute quality by tuning the optimal combination of these runners and managing the resulting massive search space of hyperparameters.

## Key takeaways

- Hedge Fund Decision Model: A hedge fund's action is conceptualized as a function: `(External Context, Internal Context) -> (Selection, Size, Timing)`. The product is a time series of this tuple, constructing a portfolio.
- Context Inputs: Inputs must be easily accessible via API. External context includes news, prices, and filings. Internal context includes the firm's current positions, liquidity requirements, and historical case studies.
- The Multi-Runner Approach: No single technology is sufficient. The optimal solution for any given step requires selecting the best 'runner': deterministic Code (fast, cheap, reliable), Statistics (for complex pattern modeling), LLMs (for summarization/state-of-the-art text understanding), or a Human (for judgment and real-world navigation).

## Technical details

- System Function Decomposition: The overall decision function is broken down into sequential, input/output-driven steps: 1) Gather (filtering relevant information), 2) Analyze (forming a market view), and 3) Decide (generating the final decision object).
- Information Gathering Complexity: The gathering step involves multiple technical choices, including data chunking, image processing (images to text), and filtration (e.g., filtering out non-material NBA news). These choices act as hyperparameters that must be tuned.
- The Search Space: The complexity is massive, involving combinatorics between the four runners (Code, Stats, LLM, Person). The system's search space is described as 'basically infinite,' requiring compartmentalization to manage the hyperparameter tuning.

## Practical implications

- Build systems should adopt a modular, pipeline-based architecture where complex processes are broken down into discrete, testable functions (Gather, Analyze, Decide).
- Design should treat the selection of processing logic (Code, LLM, Statistical Model, Human Review) as a tunable hyperparameter, prioritizing quality over the sheer number of automated steps.
- The architecture must support cross-referencing information from disparate sources and integrating multiple 'runners' (e.g., deterministic code for polling, LLM for summarization, and human review for final judgment).

## Topics

AI Agents, Financial Modeling, System Architecture, Pipeline Engineering, Hyperparameter Optimization, Natural Language Processing, Frank Long, AIE CODE summit

Source: https://www.youtube.com/watch?v=AYFcmnQrZCg
