GF Securities asserts that the core of AI-driven investment research lies in its ability to faithfully replicate the real-world investment research production chain. The strategy team has systematically laid out the construction logic and advanced optimization paths for such a system, offering the industry a complete blueprint that spans from first principles to engineering practice.

On August 28th, the Liu Chenming team at GF Securities released a report highlighting that the core workflow of fundamental analysts and fund managers involves capturing information asymmetries and forming cognitive advantages, which ultimately translate into research viewpoints or investment decisions. The design of an AI research system should adhere to this as its primary principle, focusing on two key layers: information ingestion and experiential processing. The former is tasked with enhancing the breadth, quality, and timeliness of research data, while the latter is responsible for distilling real-world research methodologies into reusable skills, implemented through AI engineering principles.

The report also flags two types of inefficient practices currently prevalent. The first is the "quant-ification" of qualitative research, which sacrifices the extraction of informational and cognitive alpha. The second is single-department attempts to build a full-chain solution in isolation, which both dilutes the benefits of specialization and increases the risk of privacy breaches. The report argues that these approaches run counter to the principles of division of labor in the modern financial industry.

Agent Foundations: Harness Engineering is the Key Variable

According to the GF Securities report, as mainstream large language models (LLMs) become increasingly homogeneous, the choice of an agent foundation is essentially a trade-off regarding the "Harness"—meaning how the external operating environment is designed to improve execution stability for complex tasks without modifying model weights. The report examines three inherent technical flaws of contemporary LLMs. The first is context degradation, where longer prompts lead to diminished response quality due to the squared computational complexity of full attention mechanisms in standard Transformers, making the effective context window often significantly smaller than its nominal size. The second is sensitivity to input perturbation, where minor, semantically insignificant changes in the prompt can cause significant output shifts, a problem exacerbated by autoregressive loops. The third is output non-determinism, where identical prompts almost never yield identical outputs due to the token sampling mechanisms in decoder-only architectures.

In terms of building the agent foundation, the report categorizes options into two types: integrated products, such as Claude Code and WorkBuddy, and open-source architectures like OpenClaw and Hermes. The former comes with vendor-tuned harnesses, offering ease of use and requiring only light optimization to handle most lightweight research tasks. The latter demands a steep learning curve due to fully customizable harnesses but can achieve a significantly higher performance ceiling with appropriate optimization, enabling absolute micro-level control over the harness and LLM queries. The report's recommended best practice is to assign specific production scenarios to open-source agents while using integrated products for project maintenance, thereby maximizing the inherent strengths of both. Open-source agents are best suited for deep research reports and stock selection strategies, while integrated products can compensate for open-source agents' shortcomings in self-diagnosis, correction, and iterative improvement.

Multi-Agent Matrix: Stepwise Loading to Resolve Context Overload

To address the core pain point of context decay, the report proposes a solution involving stepwise loading and a multi-agent matrix. The underlying logic is to decompose complex, long-horizon tasks into multiple independent "minimal high-signal" sub-modules that are executed separately, preventing irrelevant context from interfering with specific steps. Based on real secondary market research scenarios, the report designs a minimal multi-agent matrix with three distinct roles: an intern, an analyst, and an investment director. The intern is responsible for querying data via APIs or MCP interfaces, performing initial cleansing, and outputting standardized information cards. The analyst handles content creation in specific sub-fields using professional research skills, saving chapter-level conclusions to local snapshots. The investment director understands user intent, decomposes tasks, plans and schedules, and coordinates the outputs of the subordinate agents. This architecture ensures that data retrieval, cleansing, and unrelated materials do not significantly compromise the quality of the final deliverables. The report notes that this logic also applies to the workflow design of integrated agents, where task flows can be templated to explicitly require maximum task decomposition and stepwise loading to avoid context buildup.

The Skill System: Blending Analytical Flexibility with Engineering Precision

Skills are another core module emphasized in the report, defined as folders that agents can dynamically load on demand, containing instructions, scripts, and resources. The design employs a three-stage, progressive disclosure approach. Metadata remains in the context to help the agent determine whether to activate the skill. The main instruction set is mounted upon activation, serving as task routing. Project resources are then loaded on demand after routing is activated. The report illustrates the engineering value of skills with two typical cases. The first is a data gateway skill. Here, the challenge is that when a single agent is configured with multiple APIs or MCPs, tool schemas residing in the context can cause attention drift and tool-calling interference. The optimization introduces a CodeAct concept, exposing only a single function signature to the agent and managing all data needs through an adapter layer script, ensuring that data retrieval does not consume LLM attention. The second is a portfolio performance attribution skill. The issue here is that sending raw stock return data directly to an LLM can yield inconsistent quantitative results due to output non-determinism. The optimization introduces a mathematical abstraction layer, converting the attribution problem into a solution of the Barra CNE10 factor model. The LLM is only responsible for translating between natural language and structured parameters, not for actual computation, thereby eliminating output randomness.

Private RAG Knowledge Bases: The Theoretical Root of Alpha and Information Asymmetry

The report’s key conclusion regarding the information ingestion layer is that private RAG vector databases are superior to LLM-Wiki approaches and represent the theoretically optimal solution for investment research. It starts with the Efficient Market Hypothesis to argue for the necessity of private knowledge bases. Using classic frameworks from Fama and Grossman-Stiglitz, the report points out that the A-share market exhibits notable structural efficiency characteristics. For under-covered securities, such as small-cap stocks with limited analyst coverage, there is significant post-earnings-announcement drift, leading to a direct rejection of the weak-form efficiency hypothesis. Even for moderately covered securities, differences in the speed of interpreting public information can still yield excess returns. This leads to the conclusion that private information, such as surveys, reverse roadshows, and personal experience, holds substantive alpha-enhancing value for research frameworks. In terms of technical choices, the report identifies unavoidable flaws in the LLM-Wiki model. First, as the corpus scale expands, it becomes difficult to maintain a globally optimal schema granularity. Second, data proliferation continuously triggers context decay, where the effective context window shrinks to only 20% to 30% of the nominal size in complex tasks. The report’s tests show that a professional investor's daily information intake is roughly 500,000 to 700,000 characters, and after two months of sustained operation, the schema management issues of an LLM-Wiki system will become significantly worse. For optimizing RAG in investment research, the report highlights two key areas. On the data preprocessing side, both images and text can be effectively deduplicated via pure algorithms, and VLM parsing offers higher accuracy than OCR for PDFs and images. For chunking strategies, structured compression followed by full mechanical chunking proves more effective than parent-child or semantic chunking. On the user query side, the report recommends intent recognition and query rewriting mechanisms, with pre-set optimal keywords for each research direction provided to the agent. Tests show this approach outperforms any form of knowledge graph tuning.

Long-Term View: AI Will Reshape Core Skills and Sell-Side Practices

The report concludes with several forward-looking assessments from the perspective of secondary financial institutions. At the individual practitioner level, the report believes that the importance of "KYA" (Know Your AI) will continue to rise. Advanced AI workflows are poised to become a fourth essential skill for practitioners, following economic theory, financial analysis, and office software proficiency. Furthermore, a micro-level understanding of AI principles will reshape practitioners' perspectives on the medium-to-long-term logic of the AI industry. For instance, the manifestation of the Jevons paradox in multi-tier model routing on the user side may impact the Return on Invested Capital assumptions of closed-source model vendors and cloud service providers. Regarding the operational format of sell-side research institutes, the report anticipates a wave of MCP-driven transformation, allowing research opinions, data, and event updates to reach institutional clients more efficiently and broadly. The report also cautions that the primary risks to this framework include AI development and application effects falling short of expectations, disruptive changes to the underlying technology paradigm of large models, and shifts in data security and regulatory policies.