AI Economist Agent: An Agentic Framework for Model-Grounded Economic Analysis with RAG, Knowledge Graphs, and Large Language Models

Researchers have proposed an agentic framework that combines retrieval-augmented generation, knowledge graphs, and large language models to produce economic analysis grounded in explicit model computations rather than raw language-model output [1]. The framework, described as an AI economist agent, is designed to address a core limitation of large language models: while they can generate fluent economic narratives, they are not inherently constrained by economic theory or real-world data [1]. The system uses knowledge graphs that encode economic data and theory, and deploys LLM-based agents to plan the analysis, retrieve relevant evidence, select appropriate models, and generate reports [1]. Crucially, quantitative claims are not produced by the language model alone; narratives are instead grounded in explicit model-based computations and linked to retrieved evidence through the agent architecture [1]. The researchers evaluated the AI economist agent on two applications. The first involved generating economist reports on U.S. inflation persistence and Federal Reserve policy [1]. The second focused on producing bank stress-test narratives for U.S. commercial real estate refinancing stress [1]. The results demonstrated that grounding the generated reports in model computations improved their economic coherence and traceability [1]. The work builds on broader trends in retrieval-augmented generation, where external knowledge sources are used to anchor language model outputs. The paper’s abstract and associated metadata were posted on arXiv under the economics general subject area on June 18, 2026 [1]. The submission includes links to code and data repositories, though the specific contents of those repositories were not detailed in the available metadata [3][5]. While the primary paper focuses on macroeconomic and financial stability applications, the challenge of grounding machine-learning narratives in structured domain knowledge extends across disciplines. For instance, in catalysis research, scientists have explored transfer learning across datasets to improve model performance when consolidating data from varied computational methods [4]. Similarly, the United Nations has acknowledged the difficulty of tracking progress on complex, interconnected goals, noting in its 2025 Sustainable Development Goals report that only 35 percent of targets were on track or making moderate progress, with nearly half moving too slowly and 18 percent in reverse [6]. These examples illustrate the broader need for analytical frameworks that can integrate diverse evidence sources while maintaining traceability to underlying data and models.