Say goodbye to the "illusion" and "forgetfulness" of large models: Building an AI investment research system with multi-agent matrix and skill system.

GF Securities believes that the core of AI investment research lies in whether it can maximize the reproduction of the real investment research production chain. The strategy team systematically explained the construction logic and advanced optimization path of the AI investment research system, providing the industry with a complete framework from underlying principles to engineering practice.

On August 28, Liu Chenming's team at GF Securities published a report pointing out that the core workflow of subjective researchers and fund managers is to capture information gaps, form cognitive gaps, and ultimately transform them into research opinions or investment decisions.

The design of an AI-powered investment research system should take this as its first principle, focusing on two core aspects: the information intake layer and the experience processing layer. The former is responsible for improving the breadth, quality, and timeliness of investment research data, while the latter is responsible for transforming real-world investment research systems into reusable skills and implementing them with the help of AI engineering principles.

The report also specifically names two types of inefficient practices: first, the "financial engineering" of subjective investment research, which leads to the loss of the source of excess returns from extracting information and cognitive gaps; second, single-stage institutions attempting to connect the entire chain, which weakens the advantages of specialization and creates risks of privacy leaks. The report argues that both of these directions contradict the principles of specialization in the modern financial industry.

Agent Base: The Harness project is the core variable.

A report by GF Securities points out that as the capabilities of mainstream large models become more homogenized, the essence of choosing an agent platform is a trade-off in Harness— that is, how to design an external operating environment to improve the execution stability of complex tasks without modifying the model weights.

The report begins by examining three inherent technical flaws in contemporary large models.

One issue is context decay : the longer the prompt input, the worse the quality of the model's response. The computational complexity of full attention in the standard Transformer increases by O(n²) with the sequence length, and the actual effective context window is often significantly smaller than the nominal window.

Secondly, it is sensitive to input perturbations : minor semantically irrelevant changes in the Prompt can cause a significant shift in output quality due to changes in the token sequence distribution, and the autoregressive loop will continuously amplify this problem.

Thirdly, there is output nondeterminism : the output of two sends of the same prompt will almost never be the same, which is due to the token probability sampling mechanism of the Decoder-Only architecture.

Regarding path selection, the report categorizes agent platforms into two types : integrated products (such as Claude Code, WorkBuddy, etc.) and open-source architectures (such as OpenClaw, Hermes, etc.).

The former, Harness, is pre-tuned by the manufacturer, making it easy to use. With slight optimization, it can handle most lightweight research tasks . The latter, Harness , is globally customized by the user, making it extremely difficult to learn. However, with proper optimization, its performance ceiling is significantly higher, enabling absolute micro-control over Harness and LLM Query.

The report recommends the following best practices: open-source architecture agents should handle specific production scenarios, while integrated agent products should handle project maintenance, maximizing the inherent capabilities of both. The former is deeply adapted to investment research scenarios, undertaking complex and long-term tasks such as in-depth research reports and stock selection strategies; the latter fills the gaps in self-diagnosis, self-correction, and self-iteration of open-source architecture agents.

Multi-proxy matrix: Step-by-step loading to crack context backlog

To address the core pain point of context decay, the report proposes a solution of step-by-step loading and multi-agent matrix. The underlying logic is to break down complex long-term tasks into multiple independent "minimum high signal set" sub-modules, which are executed independently step by step to avoid irrelevant context from interfering with the execution effect of specific stages.

Based on real-world investment research scenarios in the secondary market, the report designed a minimal multi-agent matrix comprising three roles: intern , analyst , and investment director. The responsibilities of each role are clearly defined and layered.

Interns are responsible for calling API or MCP interfaces to query data, performing preliminary cleaning, and outputting standardized information cards;Analysts are responsible for using their professional investment research skills to create content in specific areas and saving chapter-based conclusions as local snapshots.The investment director is responsible for understanding user intent, breaking down tasks, planning and scheduling, and coordinating the content output of sub-instances.

The goal of this architecture is to ensure that data retrieval, cleaning, and removal of irrelevant information do not significantly affect the quality of the final deliverables.

The report points out that this approach also applies to the workflow design of integrated agents. When the task process is condensed into a reusable workflow, it can be explicitly required to "maximize the splitting and step-by-step loading of tasks and avoid context accumulation".

Skill System: Combining Analytical Flexibility with Engineering Precision

Skills is another core module emphasized in the GF Securities report. It refers to folders that can be dynamically loaded by the agent on demand, containing instructions, scripts, resources, etc. Its design adopts a three-stage, progressive disclosure approach :

Metadata resides in the Context for the Agent to determine whether activation is required;The main body of the instruction is mounted upon activation; its essence is task routing.Project resources are loaded on demand after the route is activated.

The report uses two typical case studies to illustrate the engineering value of Skill.

Case 1 is the investment research data gateway Skill. Its engineering pain point is that when a single agent is configured with multiple APIs or MCPs, the persistent context of the Tool Schema will cause attention drift and interference with tool calls.

The optimization solution is to introduce the CodeAct concept, which exposes only a single function signature to the Agent, and the adaptation layer script manages all data requirements in a unified manner, ensuring that the data retrieval process does not consume the attention of the LLM.

Case 2 is the portfolio performance attribution skill . Its engineering pain point is that sending individual stock price change data directly to LLM analysis will result in random differences in quantitative results due to the uncertainty of model output.

The optimization scheme introduces a mathematical abstraction layer, which transforms the attribution problem into a mathematical solution for the Barra CNE10 factor model. The LLM is only responsible for translating between natural language and structured parameters and does not participate in the actual calculation, thereby eliminating the randomness of the output.

Private Domain RAG Knowledge Base: Information Gap is the Theoretical Root of Alpha

The report's core conclusion at the information ingestion level is that the private domain RAG vector knowledge base is superior to LLM-Wiki and is the theoretically optimal solution for investment research scenarios.

The report begins by arguing for the necessity of private knowledge bases based on the efficient market hypothesis (EMH).

Based on classic frameworks such as Fama and Grossman-Stiglitz, the report points out that the A-share market exhibits obvious structural efficiency characteristics: under-covered targets (small market capitalization, 0-2 or fewer analysts) show significant post-announcement price drift (PEAD), directly rejecting the weak-form efficiency hypothesis.

For medium-coverage targets, differences in the speed at which publicly available information is interpreted can still generate excess returns. Therefore, it can be inferred that private domain information (research, reverse roadshows, personal experience, etc.) has substantial alpha-enhancing value for the investment research framework.

Regarding path selection, the report points out that LLM-Wiki has an inherent, unsolvable flaw. First, as the corpus grows, it becomes difficult to maintain the globally optimal schema granularity; second, data expansion continuously triggers context decay, and in complex tasks, the effective context window of LLM is only 20% to 30% of the nominal size.

The report's actual measurements show that professional investors consume approximately 500,000 to 700,000 words of qualified information per day, and the schema management issues of LLM-Wiki will significantly worsen after more than two months of continuous operation.

Regarding the optimization of RAG's investment research scenarios, the report proposes two key aspects.

On the data preprocessing side , both images and text can be efficiently deduplicated using pure algorithms. PDFs and images can be parsed with VLM at a higher accuracy than OCR solutions. In terms of slicing strategies, the effect of complete mechanical slicing after structured compression is better than special solutions such as parent-child slicing and semantic slicing.

On the user query side , the report suggests using intent recognition and query rewriting mechanisms, and pre-setting optimal keywords for each research direction for the agent to complete; actual tests show that this approach is more effective than any form of knowledge graph optimization.

Long-term thinking: AI capabilities will reshape the basic skills of practitioners and the way sellers conduct business.

The report concludes with several forward-looking judgments from the perspective of secondary financial institutions.

In terms of practitioners' capabilities, the report believes that the importance of "KYA (Know Your AI)" will continue to rise—advanced AI workflows are expected to become the fourth essential basic skill for practitioners, following economic thinking, financial analysis, and office software operation.

At the same time, a micro-level understanding of AI principles will reshape practitioners’ perception of the long-term logic of the AI industry . For example, the Jevons paradox in the context of multi-level model routing on the user side may affect the ROIC assumption of closed-source model vendors and cloud service providers (CSPs).

Regarding the business model of sell-side research institutes, the report predicts that the industry may undergo a round of MCP-like transformation, where research viewpoints, data, and activity updates will be able to reach institutional clients more efficiently and broadly.

The report also warns that the main sources of risk for the aforementioned framework are the underperformance of AI technology development and application, disruptive changes in the underlying technological paradigm of large-scale models, and changes in data security and regulatory policies.

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.