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

AI agents will have an even bigger impact on investing than they have on coding, but realizing that impact requires decomposing the investment decision proce

AI Agents and the Decomposition of Investment Decisions: A Pipeline Framework from Coatue

Abstract

This synthesis examines a framework, articulated by Frank Long of Coatue, for applying AI agents to discretionary investment management. The central claim is that agents will have a larger effect on investing than on software engineering, provided the investment process is decomposed into granular steps rather than delegated wholesale to a model. The framework models a hedge fund as a function mapping external and internal context to a decision tuple - selection, size, timing - then recursively decomposes that function into a gather-analyze-decide pipeline whose nodes are assigned to one of four "runners": code, statistics, LLMs, or humans. A worked example involving Korean-language semiconductor news shows that each pipeline node carries roughly ten hyperparameters with twenty or more valid values, producing an effectively infinite configuration space that requires an explicit optimization objective to navigate productively.

1. Introduction

Discretionary investing is often treated as resistant to automation because it depends on judgment. The framework examined here reframes that assumption: judgment is a property of specific nodes within the investment process, not a property of the process as a whole, and most nodes are not judgment-bearing. The thesis is stated directly:

"I think that agents will have an even bigger impact in investing than they did at coding."

This claim is paired with an explicit caveat - realizing the impact requires engineering discipline, not a larger context window. Key terminology anchors the analysis. External context refers to information originating outside the firm, such as news, prices, and regulatory filings. Internal context refers to firm-specific state: current positions, liquidity and risk constraints on particular names, and institutional history with a sector, company, or management team. A runner denotes the execution modality - code, statistics, an LLM, or a human - assigned to a given step in the decomposed process.

The analysis proceeds in three stages: defining the hedge fund as a compressible function, decomposing that function into a gather-analyze-decide pipeline, and assigning runners to each resulting node according to a quality-maximization criterion rather than an automation-maximization criterion. A case study on Korean-language news illustrates the combinatorial complexity that emerges even from a single data source.

2. Background and Related Work

The framework borrows its core methodological apparatus from supervised machine learning, specifically hyperparameter tuning. Where conventional ML practice tunes learning rates or layer widths, this approach treats pipeline design decisions - chunking strategy, image-versus-text processing, translation policy, model size, reasoning effort - as hyperparameters subject to search and optimization. This analogy supplies both a vocabulary for the design space and a requirement that search be directed by an explicit objective, described as a "North Star," rather than conducted by ad hoc experimentation.

A second prerequisite identified is organizational rather than technical. The firm had for decades maintained written notes on every trade and every decision made or declined, recording outcomes. This archive constitutes internal context that most institutions lack and is credited with positioning the firm for AI adoption. KKR is cited as a leader in building this kind of foundation over more than a decade, both technologically and culturally, indicating that data infrastructure alone is insufficient without the institutional practices that populate it.

3. Core Analysis

3.1 The Investment Process as a Compressible Function

A hedge fund is modeled as a function accepting two parameters - external and internal context - and returning a tuple: selection, size, timing. The firm's output is a time series of such tuples that jointly constructs a portfolio. Framed this way, the enterprise is fundamentally a compression problem: all available information in the world must be compressed into three decision variables. This framing motivates the claim that agents matter more here than in coding, since coding already operates within a relatively compressed, syntactically constrained space, whereas investment decisions compress an open-ended information environment.

A supporting principle governs the input side of this function: "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." The system's intelligence is bounded by the tools it can action, analogized to hiring expert investors but denying them logins to most of the firm's services. External context (news, prices, filings) and internal context (positions, risk constraints, historical notes) must both be API-accessible rather than retrieved manually.

3.2 Decomposition into Gather, Analyze, Decide

Because the function cannot be collapsed directly into a single model call, it is decomposed into three sequential steps, each with a defined input and output, where the output of one step becomes the input of the next:

1. Gather - collect external and internal context and filter material from immaterial information (e.g., NBA news is immaterial; local permitting news may be material to data-center exposure).
2. Analyze - form a view on what gathered information implies for public and private markets.
3. Decide - convert that view into the decision tuple (selection, size, timing).

This decomposition is itself recursive: each of the three steps further decomposes into sub-steps, each of which requires its own runner assignment and hyperparameter choices.

3.3 Case Study: Korean-Language Semiconductor News

Korean- and broader Asian-language media is identified as critical during a period when semiconductors and AI dominate market themes, and AI models are positioned to process Korean-language sources more effectively than a non-Korean-speaking staff. Decomposing just this single data source yields multiple sub-steps: article collection (via MCP, CLI, or licensed and cleaned data feeds), chunking decisions, and a choice between processing document images directly or converting them to text. A filtration step removes immaterial content (Korean NBA news) from content relevant to portfolio names and themes. Subsequent sub-steps include translate/extract/summarize decisions, structuring extracted information, and cross-referencing it against internal notes, prior views, conversations, and meeting recordings. This cross-referencing across disparate sources is characterized as especially valuable. Each of these choices - collection method, chunking, image handling, filtration threshold, summarization approach - is treated as a tunable hyperparameter, directly extending the ML-tuning analogy from Section 2 to pipeline architecture itself.

3.4 The Four Runners and the Quality Criterion

Each node in the decomposed pipeline is assigned to one of four runner categories: code, statistics/classical ML, LLMs/AI agents, or people. Code is favored for fast, cheap, deterministic rules (e.g., "if price-to-earnings ratio is less than 15"). Statistics or classical ML is favored where relationships are more complex than fixed rules but still tractable via regression or similar fitting. LLMs are favored for tasks like summarization that resist if-statement formalization. People remain the preferred runner where judgment and real-world navigation are required, such as meeting with a management team - a task explicitly excluded from LLM or agent delegation.

The governing principle is stated plainly: "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 optimization target is per-node quality, not aggregate automation: "The goal is, for any given piece of this function, what can I do to maximize absolute quality?" Most production steps are consequently multi-runner, combining code, statistics, LLMs, and humans within a single node rather than relying on one modality.

4. Technical Insights

Within the LLM runner category alone, additional hyperparameters stack on top of pipeline-level choices: model size (big/small/medium), reasoning effort (extra high/high/medium), harness selection, which skills load with the harness, and which hooks load with the skills. Combined with pipeline-level hyperparameters (chunking, filtration, translation policy, source-collection method), each node in the overall dependency graph is estimated to carry at least ten hyperparameters with at least twenty valid values each - a search space described as effectively infinite. This combinatorial structure implies that without an explicit optimization objective, configuration effort degenerates into undirected fidgeting rather than systematic improvement. A practical implication is that pipeline design decisions should be logged and evaluated with the same rigor as model hyperparameters in conventional ML workflows, including version control over configuration choices and measurable quality criteria per node.

5. Discussion

The framework's broader significance lies in reframing "AI in investing" as a systems-engineering problem rather than a model-capability problem. The constraint is not whether an LLM can summarize Korean news adequately - it largely can - but whether the surrounding pipeline (collection, filtration, cross-referencing, runner assignment) is architected to make that capability useful at scale. This mirrors trends in applied AI more broadly, where production value increasingly depends on orchestration and evaluation infrastructure rather than raw model capability.

A notable gap is the absence of a specified methodology for defining the "North Star" objective that governs hyperparameter search at each node - the framework asserts its necessity without detailing how node-level quality is measured or validated against realized investment outcomes. The internal dependency graph, referred to as Kensho, is acknowledged as "massively complex," and the primary challenge is described as organizational and cultural comprehensibility as much as technical tractability - an admission that technical decomposition alone does not guarantee adoption.

The framework's authors note its applicability beyond investing, to policy, nonprofits, drug discovery, and corporate strategy - domains sharing the structure of compressing large, heterogeneous information environments into discrete, consequential decisions.

6. Conclusion

This analysis contributes a decomposition methodology - gather, analyze, decide, with four-runner assignment - that reframes judgment-dependent decision processes as graphs of mostly non-judgment nodes, each independently optimizable. The practical takeaway is that effective AI deployment in investing does not require maximizing LLM usage but instead requires rigorous, quality-driven assignment of each pipeline node to its most suitable runner, supported by API-level access to all relevant context and institutional history. Organizations pursuing similar AI integration should prioritize building the dependency graph and defining per-node quality objectives before expanding LLM-based automation.

Sources

• AI Agents Will Upend Investing Even More Than Coding - Frank Long, Coatue - Original Creator (YouTube)
• Analysis and summary by Sean Weldon using AI-assisted research tools

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub