When Analysts Become Agents: Fine-Grained AI Teams That Actually Trade

Trading teams rarely fail because nobody had a title.

They fail because the signal gets lost somewhere between the analyst, the sector specialist, the portfolio manager, and the final trade list. Someone sees momentum. Someone else sees valuation. A news analyst notices a red flag. A macro analyst says the regime is awkward. Then the PM receives a pile of half-compatible opinions and performs the ancient institutional ritual known as “synthesis,” which is often just a polite word for discretionary compression.

A new paper, Toward Expert Investment Teams: A Multi-Agent LLM System with Fine-Grained Trading Tasks, asks a useful question: if we build LLM trading agents to imitate investment teams, should we merely assign them roles, or should we also encode the actual work process behind those roles?¹

That distinction sounds small. It is not. “You are a technical analyst; analyze prices” is not the same as “evaluate one-month long-short attractiveness using normalized momentum, Bollinger deviation, MACD, RSI, and stochastic indicators, then output a bounded score and short rationale.” The first prompt gives the model a job title. The second gives it a workflow.

The paper’s answer is refreshingly inconvenient for the current agent hype cycle: more agents are not automatically better, and more raw data is not automatically smarter. The useful unit is not the agent. The useful unit is the task that survives the handoff.

The experiment is not about magical AI traders

The study builds a hierarchical LLM investment team for Japanese equities. The universe is TOPIX 100 stocks. The strategy is monthly rebalanced, market-neutral, and long-short: the system ranks stocks, buys the highest-scored names, shorts the lowest-scored names, and rebalances at the opening of the first business day of each month.

The architecture resembles a simplified institutional investment workflow:

The model used is GPT-4o. The backtest runs from September 2023 to November 2025, a period chosen to sit after the model’s August 2023 knowledge cutoff. The authors also feed agents only information available up to each decision point. That does not make the backtest immune to every possible form of bias, but it does address the obvious “the model already knew the future” problem.

The core comparison is between two prompt designs for the Technical and Quantitative Agents:

This is not a contest between “LLM with data” and “LLM without data.” Both sides receive data. The difference is whether the model must rediscover the analyst workflow from raw inputs, or whether the workflow is already shaped into investment tasks.

That is the first business lesson hiding in the paper. The system does not become more professional because an agent has the word “analyst” in its system prompt. It becomes more professional when the task resembles something an analyst would actually do.

The main evidence: fine-grained tasks win in most portfolio sizes

The paper first compares fine-grained and coarse-grained task settings using all agents. It evaluates five portfolio sizes: 10, 20, 30, 40, and 50 total positions, where the long and short sides are balanced.

Across 50 independent trials, the fine-grained configuration significantly outperforms the coarse-grained one in 4 out of 5 portfolio sizes: 20, 30, 40, and 50. The 10-stock portfolio is the exception and is not statistically significant.

That exception matters. A narrow long-short portfolio can be noisy because a few stock selections dominate outcomes. The fact that the result strengthens in broader portfolios suggests the paper is not merely showing one lucky concentrated bet. It is closer to saying: when enough names are included for the ranking system to express itself, the task decomposition matters.

The leave-one-out comparison sharpens the finding. The authors remove one analytical component at a time and compare the median Sharpe ratio difference between fine-grained and coarse-grained designs.

The important row is “Without Technical.” In the all-agent system, fine-grained prompts win in most settings. But once the Technical Agent is removed, the fine-grained advantage collapses and turns negative for larger portfolios.

This is the paper’s first real twist. The result is not simply “fine-grained prompts are better.” It is more specific: fine-grained prompts seem especially useful when they help technical signals become usable inside the hierarchy.

That is a narrower claim, and therefore a better one.

The Technical Agent is the uncomfortable star

In polite finance conversations, technical analysis often sits in a strange social position. Many practitioners use price and momentum signals. Many also prefer not to sound too enthusiastic about them at conferences where everyone is pretending to be a discounted-cash-flow philosopher.

This paper does not settle the philosophy. It does show that, in this specific system, the Technical Agent is doing a lot of the work.

The ablation study compares each configuration with the all-agent baseline. Here the sign convention is important: a positive number means removing the agent improved the Sharpe ratio; a negative number means removing the agent hurt performance.

In the fine-grained setting, the all-agent baseline Sharpe ratios are 0.54, 0.84, 0.84, 0.79, and 0.90 across the five portfolio sizes. Removing the Technical Agent produces negative differences in four of the five sizes, especially in wider portfolios: -0.42, -0.40, -0.56, and -0.66 for 20 through 50 positions.

In plain English: with fine-grained task design, the Technical Agent is valuable. Remove it, and the system gets worse.

Other agents behave differently. Removing the Quantitative, Qualitative, News, or Macro Agents often improves Sharpe in the fine-grained setting. That does not mean those information sources are useless in real investing. It means that, inside this particular LLM hierarchy, some agents may add redundant, weak, or noisy signals once the technical workflow is already doing useful ranking work.

That is a sharp warning for agent-product builders. A multi-agent system can look more sophisticated while becoming less effective. The extra agent creates another place for signal dilution, another textual handoff, another opportunity for the PM agent to over-average conflicting views. Bureaucracy has finally been automated. Congratulations, I suppose.

The coarse-grained setting tells a different story. There, removing the News Agent hurts performance across most portfolio sizes, while the Technical Agent’s contribution is weaker and less consistent. The authors interpret this as possible evidence that, without fine-grained technical decomposition, news may compensate for weaker propagation of technical information.

This is not a universal law about news versus technicals. It is a system-level observation: when one channel is poorly structured, another channel may become relatively more important. In agent design, the marginal value of an information source depends on how the rest of the workflow is shaped.

The paper’s interpretability tests are not decoration

The most useful part of the paper is not merely the backtest table. It is the attempt to inspect intermediate outputs.

The authors run two kinds of text analysis. First, they use log-odds analysis to compare representative words in fine-grained versus coarse-grained agent outputs. Second, they use embeddings to measure semantic similarity between lower-level agent outputs and higher-level outputs, especially the Sector Agent.

The log-odds analysis shows a vocabulary shift. Fine-grained technical outputs are associated with words such as “Momentum,” “Neutral,” “Short-term,” “Favorable,” “Condition,” “Decline,” and “Volatility.” Coarse-grained technical outputs lean more toward broad price-language: “Price,” “Trend,” “Rise,” “Upward-trend,” “Fall,” “Stock-price,” and “Continuation.”

The quantitative and higher-level agents show a similar pattern. Fine-grained outputs contain more ratio-like and analytical vocabulary: “Soundness,” “Margins,” “Profitability,” “Growth-rate,” “ROE,” “Efficiency,” and “Undervalued.” Coarse-grained outputs contain more generic financial-statement terms such as “EPS,” “Profit,” “Operating-profit,” “Net-income,” and “Average.”

The tempting interpretation is “fine-grained prompts make the model sound smarter.” That is possible, and it is not enough. A better interpretation is that fine-grained prompts make the agent’s language more compatible with the downstream decision process.

The embedding analysis supports this more operational reading. The Sector Agent’s output is naturally more similar to fundamental and qualitative agents than to the Technical Agent. In the reported median cosine similarities, the Qualitative Agent has the highest similarity with the Sector Agent at 0.514 in both settings. The Quantitative Agent is also relatively high, around 0.476. The Technical Agent is lower: 0.419 in the fine-grained setting and 0.397 in the coarse-grained setting.

The absolute value is not the story. The change is the story. Technical-to-Sector similarity increases by 0.022 under the fine-grained setting, while Quantitative and Qualitative similarity barely changes.

This is where the paper becomes more interesting than a “prompt engineering improves Sharpe” headline. The fine-grained technical prompt does not merely improve the Technical Agent in isolation. It appears to help technical reasoning survive the transition into the Sector Agent’s synthesis.

In a hierarchical agent system, that is the whole game. A signal that does not propagate is not a signal. It is a memo nobody reads.

The portfolio test is about diversification, not replacement

The final experiment examines portfolio optimization. This section is easy to misread, so let us be precise.

The authors create a composite of six LLM-based agent strategies: the all-agent strategy and five leave-one-out strategies. They combine those agent strategies using an equal-risk-contribution approach, then test allocations between this agent composite and the TOPIX 100 index. The agent strategy includes a one-way transaction cost assumption of 10 basis points.

The results are:

A careless reading would say the AI agent beats the market. It does not, at least not as a standalone sleeve in this table. The TOPIX 100 has a higher net Sharpe than the agent strategy alone: 1.68 versus 0.95.

The stronger result is subtler. The 50-50 combined portfolio has lower volatility and a higher Sharpe than either standalone component. Net of transaction costs, the combined portfolio produces a Sharpe ratio of 1.91, above TOPIX 100 alone and above the agent composite alone.

This is exactly how many institutional innovations become useful. They do not need to replace the core portfolio. They need to add a return stream that is sufficiently different from the core portfolio. In this paper, the agent composite is valuable less because it dominates TOPIX and more because it diversifies it.

That is a more credible business pathway. “Deploy autonomous AI trader with live capital” is a governance headache with a marketing department attached. “Test an LLM-derived signal sleeve as a diversifying overlay under risk controls” is still difficult, but at least it resembles something an investment committee might discuss without everyone pretending they did not hear it.

What the evidence supports, and what it does not

The paper’s evidence is layered. Treating every table as the same kind of proof would flatten the argument. A better reading is:

This distinction matters because business readers often compress research into one takeaway. Here the defensible takeaway is not “LLM agents can trade.” It is “LLM agent systems become more testable and sometimes more effective when expert workflows are decomposed into concrete tasks and when intermediate outputs are monitored.”

Less exciting? Perhaps. More useful? Unfortunately, yes.

The business lesson is workflow engineering, not agent theater

For AI investment-product builders, the paper points toward a concrete operating model.

First, turn analyst SOPs into explicit prompt protocols. Do not ask a “fundamental agent” to analyze fundamentals in the abstract. Define the ratios, horizons, missing-data rules, score scales, and output contracts. The more expensive the downstream decision, the less charming ambiguity becomes.

Second, separate information sources from analytical tasks. A data modality is not a task. “Financial statements” are a source. “Evaluate profitability trend using TTM margins and compare current ROE with sector context” is closer to a task. The paper’s results suggest that this distinction can affect not just output style but portfolio behavior.

Third, monitor intermediate outputs. The paper’s text analysis is not merely academic garnish. In enterprise deployment, intermediate agent text is an audit surface. If a technical signal drives performance, one should be able to see whether it reaches the sector layer and portfolio layer. If it disappears, the problem may not be data quality. It may be handoff design.

Fourth, test marginal agent value. Adding a Macro Agent, News Agent, or Qualitative Agent sounds prudent. The ablations show that prudence can become noise. Each agent should earn its place through marginal contribution tests, not through organizational aesthetics.

Fifth, validate agent outputs as overlays before treating them as autonomous portfolios. The portfolio-optimization result is a sensible deployment metaphor. An LLM strategy sleeve can be evaluated by correlation, volatility, turnover, transaction cost, and marginal contribution to a broader portfolio. This is more boring than a fully autonomous robo-PM. Boring is underrated when capital is involved.

Where the result should not be over-sold

The boundaries are real.

The backtest covers 27 months, from September 2023 to November 2025. That is intentionally chosen to reduce look-ahead concerns relative to GPT-4o’s August 2023 knowledge cutoff, but it is still a short market history. Japan over that period is not every market regime.

The universe is TOPIX 100. Large-cap Japanese equities have their own disclosure patterns, liquidity, sector structure, and macro sensitivities. A design that works there may behave differently in US equities, crypto, small caps, emerging markets, or intraday futures. The paper itself notes the need to test other markets and models.

The model is GPT-4o, using a specific prompt architecture and Japanese-language outputs. Some of the observed gains may come from vocabulary patterns that downstream agents happen to adopt more readily. The authors explicitly acknowledge this possibility. In other words, “fine-grained task decomposition” and “LLM preference for certain linguistic cues” may be entangled.

The transaction-cost assumption appears in the portfolio-optimization section, but trading frictions are always strategy-specific. Monthly TOPIX 100 rebalancing is not high-frequency trading, but real-world deployment would still need careful treatment of borrow costs, shorting constraints, market impact, execution timing, taxes, capacity, and operational controls.

Most importantly, the paper does not prove that a multi-agent LLM system should be trusted with autonomous capital allocation. It shows that a particular design improves risk-adjusted backtest performance over a coarse-grained prompt baseline, that technical signal propagation appears central, and that an agent composite can improve a combined portfolio’s Sharpe ratio in this test setting.

That is enough to be interesting. It is not enough to fire the portfolio manager. The machines will have to suffer through committee meetings like everyone else.

From role prompts to operating systems

The agent industry likes job titles. Researcher Agent. Analyst Agent. PM Agent. Risk Agent. The names are comforting because they map machine behavior onto familiar organizations.

This paper suggests that the name is the least important part.

A trading agent becomes useful when its task is specific enough to generate stable, inspectable, and downstream-compatible outputs. A multi-agent system becomes useful when signals survive the hierarchy. And an investment product becomes credible when the output is tested not only as a standalone miracle machine, but as a component inside a portfolio construction process.

The best phrase in the paper is not a slogan. It is the implied design principle: task granularity is infrastructure.

That principle travels beyond trading. Customer-support agents, procurement agents, legal-review agents, financial-reporting agents, and research agents all face the same basic problem. A vague role prompt asks the model to invent the workflow. A fine-grained task protocol gives it the workflow and lets the model operate inside it.

The future of enterprise agents may not belong to the biggest artificial org chart. It may belong to the team that knows exactly what each agent is supposed to do, what it must pass forward, and how to tell when it is adding noise.

A shocking discovery: management still matters.

Cognaptus: Automate the Present, Incubate the Future.