AI Agents for Economic Research | NBER

The paper "AI Agents for Economic Research" by Anton Korinek, published as an NBER Working Paper in September 2025, aims to demystify autonomous AI agents—systems based on large language models (LLMs) that can plan, use tools, and execute multi-step research tasks—and provide practical guidance for economists to build their own, even without programming expertise. The work illustrates how AI has evolved from simple chatbots to reasoning models and now to autonomous agents capable of transforming economic research workflows.

Through step-by-step code examples and conceptual frameworks, the paper demonstrates how economists can create AI agents that autonomously conduct literature reviews across vast sources, write and debug econometric code, fetch and analyze economic data, and coordinate complex research workflows. A key innovation highlighted is "vibe coding," or programming through natural language, which allows researchers to build sophisticated research assistants quickly using modern agentic frameworks like LangGraph.

LangGraph enables the creation of customizable multi-agent systems by allowing researchers to define task nodes via natural language descriptions—such as “download data from FRED,” “run regression,” or “write results summary”—and connect them into executable workflows. These agents can autonomously determine subsequent actions based on intermediate outputs, exhibiting genuine autonomy and enabling the development of a multi-agent ecosystem for research.

The paper also presents complete working implementations of such agents, emphasizing their integration into every stage of the research process, from initial investigation to final analysis. It builds upon earlier work, including Korinek’s 2023 and 2024 updates on generative AI in economics, showing how AI tools have rapidly matured from conceptual prototypes into practical, daily-use systems within just two years.

Furthermore, the guide addresses the emergence of agentic chatbots and multi-agent architectures that can synthesize hundreds of academic sources in minutes, producing comprehensive research reports with accurate citations. As standardized protocols like MCP (Model Context Protocol) and A2A (Agent-to-Agent) evolve, seamless collaboration between AI agents becomes increasingly feasible, signaling a structural transformation in how economic research is organized and conducted.

The practical implications are significant: these tools not only enhance productivity but also democratize access to advanced computational methods, potentially leveling the technical playing field in the discipline. Economists are encouraged to engage with these systems actively to understand their capabilities and limitations, with the paper providing immediate, modifiable code examples to support experimentation and adaptation.

This working paper serves as a comprehensive introduction to the architecture and application of autonomous LLM-based agents within the context of economic research. Moving beyond simple prompt-and-response interactions, the authors define AI agents as systems that utilize Large Language Models as central controllers to plan, execute, and debug multi-step workflows. The material covers the technical infrastructure required to build these agents, focusing on how LLMs can be integrated with external tools—such as code interpreters, file systems, and web browsers—to perform complex tasks like data collection, cleaning, and analysis. It effectively demystifies the "agentic" loop, where the model iteratively observes the state of a task, reasons about the next action, and executes code until a research objective is met.

A key contribution of the paper is its practical, hands-on approach to agent design, providing economists with the blueprint to deploy these systems using frameworks like LangChain or AutoGen. The authors detail specific prompting strategies and architectural patterns (e.g., multi-agent debates or planner-executor hierarchies) that enhance performance and reduce hallucination in quantitative reasoning. Furthermore, the paper offers a critical assessment of current capabilities, benchmarking agents against traditional research workflows. It highlights that while current models can autonomously replicate sections of existing empirical papers and handle sophisticated coding tasks, they still require significant human oversight for valid causal inference and economic interpretation.

The significance of this work lies in its potential to radically reshape the productivity frontier of economic research. By automating the "grunt work" of data wrangling and coding, these agents allow researchers to focus on higher-level conceptualization and theory. The paper acts as a foundational guide for the transition from using AI as a sophisticated chatbot to using it as a research collaborator, signaling a shift toward "AI-native" economic science. It also prompts necessary discussions regarding reproducibility, verification, and the changing skill set required for economists in an era of increasingly autonomous artificial intelligence.

# AI Agents for Economic Research (NBER Working Paper w34202)

This paper introduces a framework for leveraging autonomous large language model (LLM)-based AI agents to automate complex, multi-step economic research tasks. It demystifies the architecture of these agents—comprising LLMs, retrieval-augmented generation (RAG), tool use, and memory systems—and provides practical guidelines for economists to implement them. The authors demonstrate how these agents can perform tasks such as literature reviews, data analysis, and even hypothesis testing, reducing the manual burden on researchers while maintaining interpretability and reproducibility.

The key contributions include: - A modular design for AI agents that integrates LLMs with domain-specific tools (e.g., Stata, Python, or economic databases). - Benchmarking of agent performance on structured tasks, showing comparable accuracy to human researchers in some cases. - Ethical and practical considerations, such as hallucination mitigation, cost management, and auditability.

This work matters because it bridges the gap between AI and economics, offering a scalable solution for tedious research tasks while preserving the rigor required in economic analysis. As LLMs advance, such agents could revolutionize how economists conduct empirical work, particularly in areas requiring iterative exploration of large datasets or interdisciplinary synthesis. For technically minded readers, the paper serves as both a theoretical foundation and a practical blueprint for deploying AI assistants in research workflows.