AI Agents for Economic Research: How Agentic AI Is Transforming the Discipline

The Rise of AI Agents in Economic Research

The field of economics is undergoing its most significant methodological transformation since the advent of computational statistics. AI agents for economic research — autonomous systems that combine large language models with reasoning, tool access, and planning capabilities — are rapidly moving from experimental curiosities to indispensable research infrastructure. A landmark NBER working paper by Anton Korinek (Working Paper 34202, September 2025) provides the most comprehensive analysis to date of how these agents are reshaping every stage of the research process, from ideation and literature review to data analysis and publication.

What makes this moment different from previous waves of AI enthusiasm is the convergence of three capabilities that, until recently, existed in isolation: natural language understanding sophisticated enough to parse economic theory, reasoning engines powerful enough to solve graduate-level mathematics, and agentic architectures that can autonomously plan and execute multi-step research workflows. The result is a new class of tools that do not merely assist economists — they actively participate in the research process. For anyone working in the social sciences, understanding this shift is no longer optional. As explored in our interactive library, the implications extend far beyond academia into policy, finance, and organizational decision-making.

This article provides a comprehensive guide to AI agents in economic research, drawing on the NBER paper’s findings, real-world benchmarks, and practical implementation details. Whether you are a graduate student exploring your first coding agent or a senior researcher evaluating whether to invest in custom agent infrastructure, this analysis will help you navigate the rapidly evolving landscape of agentic AI for economics.

From Chatbots to AI Agents: Three Paradigms of Intelligence

Understanding AI agents for economic research requires grasping the three distinct paradigms that have emerged since November 2022. Each paradigm builds on the previous one, and each offers different capabilities for different types of economic work. The NBER paper maps these paradigms directly to seven core research activities, providing economists with a practical framework for tool selection.

The first paradigm — traditional LLM-based chatbots — operates as “System 1” thinking: fast, intuitive, pattern-recognition-based text generation. These tools excel at drafting, summarizing, editing, and translating academic text. By mid-2025, performance differences between leading models had narrowed considerably. On the LMSYS Arena leaderboard, the gap between the top-ranked model (Google Gemini 2.5 Pro at 1457) and the sixth-ranked (Kimi K2 at 1421) translates to only 60/40 winning probabilities in head-to-head matchups. For basic writing and summarization tasks, model choice has become largely a matter of preference and pricing.

The second paradigm — reasoning models — emerged in September 2024 and brought “System 2” capabilities: deliberate, step-by-step problem solving trained via reinforcement learning. These models can solve multi-step mathematical proofs, write and debug complex code, and construct formal arguments at or exceeding human expert performance. On the GPQA benchmark, where PhD experts score approximately 65%, GPT-5 now achieves 89.4% and Grok-4 reaches 88.9%. By July 2025, both OpenAI and DeepMind models achieved gold-medal performance at the International Mathematical Olympiad. For economists doing formal theory work, this represents a genuine capability leap.

The third paradigm — agentic chatbots — launched in December 2024 and synthesizes language generation, reasoning, and autonomous action. These systems can plan sequences of operations, actively gather information from external sources, execute code, and adapt strategies based on what they discover. The NBER paper demonstrates this with a striking example: given a single prompt to analyze the Beveridge curve, ChatGPT o3 autonomously fetched JOLTS vacancy data and BLS unemployment data, produced a scatter plot, and delivered a detailed phase-by-phase analysis of 25 years of U.S. labor market dynamics — all without additional human input.

How AI Agents Transform Economic Research Workflows

AI agents in economic research do not simply accelerate existing workflows — they fundamentally restructure them. The NBER paper presents a systematic mapping of which AI paradigm is most productive for each of seven core research categories, and the results challenge common assumptions about where agents add the most value.

For ideation and feedback, reasoning models outperform traditional LLMs and agentic systems alike. The deliberate, step-by-step thinking process enables more rigorous stress-testing of hypotheses, identification of logical gaps, and exploration of alternative frameworks. Economists can present a research question and receive structured feedback that identifies potential identification issues, suggests robustness checks, and proposes alternative empirical strategies — all at the quality level of a well-informed colleague.

For background research and data analysis, agentic chatbots dominate. Their ability to autonomously search multiple databases, cross-reference findings, and synthesize information from dozens of sources makes them uniquely suited to the exploratory phases of economic research. Deep Research agents from Google and OpenAI can consult over 100 sources in a single query, producing comprehensive literature reviews that would take a human researcher days or weeks. The Federal Reserve Economic Data (FRED) API, combined with agent frameworks, enables automated data retrieval and preliminary analysis that previously required specialized programming skills.

For writing, traditional LLMs remain the most efficient choice. The agentic overhead — planning, tool calling, verification — adds latency without proportional quality improvement for straightforward prose generation. However, agents shine when writing requires embedded data analysis, citation verification, or multi-source synthesis.

For mathematical modeling and coding, reasoning models paired with coding agents provide the strongest combination. The paper documents that vibe coding — building software through natural language descriptions — has reached a maturity where economists can create functional econometric tools without writing a single line of code. A complete OLS regression application with CSV upload, variable selection, and visualization was built in under seven minutes through conversational prompts to Claude Code.

Deep Research AI Agents for Literature Review

One of the most immediately impactful applications of AI agents in economic research is automated deep research — comprehensive, multi-source literature reviews and environmental scans that previously consumed weeks of researcher time. The NBER paper examines both commercial and custom-built approaches, revealing significant differences in capability, cost, and reliability.

Commercial deep research systems from Google (Gemini Deep Research) and OpenAI (ChatGPT Deep Research) represent the most accessible entry point. These systems operate by breaking a research question into subtasks, executing dozens to hundreds of web searches, analyzing the retrieved content, and synthesizing findings into structured reports. In testing, Gemini Deep Research consulted over 100 sources for a single query and delivered results in 5-10 minutes. OpenAI’s system performed 109 searches narrowed to 21 curated sources, taking 5-30 minutes depending on complexity. The estimated compute cost per report is approximately one dollar.

For economists, these systems work best when investigating topics with substantial existing literature. The paper notes an important limitation: deep research AI agents struggle with frontier research topics where published work is sparse. When queried about emerging areas, they tend to over-rely on blog posts, preprints, and secondary sources rather than acknowledging the absence of established research. This makes human verification particularly important for cutting-edge topics.

Custom-built deep research agents offer superior control and dramatically lower costs. The NBER paper includes a complete implementation — approximately 370 lines of Python using the LangGraph framework — that coordinates multiple specialized sub-agents: a lead researcher for strategic planning, parallel search agents for web retrieval, analysis agents that evaluate results from different perspectives, and a synthesis agent that integrates findings into a coherent narrative. This custom system performed 15 searches across 5 subtasks in approximately one minute, at a cost of roughly one cent in API tokens. The 100x cost reduction compared to commercial alternatives makes it viable for high-volume research workflows, such as systematic literature reviews covering hundreds of papers.

A critical architectural insight from the paper is the tiered model approach: using expensive, high-capability models (like GPT-4.1) for strategic decisions and synthesis, while deploying cheaper models (like GPT-4.1-mini) for routine analysis tasks. This optimization pattern, applicable across all agent architectures, allows researchers to balance quality against cost in a principled way. For those interested in how document transformation can enhance research communication, explore our collection of research paper analyses.

Coding AI Agents and the Vibe Coding Revolution

Perhaps the most democratizing development documented in the NBER paper is the emergence of coding AI agents that enable “vibe coding” — programming through natural language descriptions rather than traditional code writing. For economic researchers who have historically been constrained by their programming abilities, this represents a paradigm shift in what is technically possible.

The paper demonstrates vibe coding through a detailed example: building a complete OLS regression tool from scratch using Claude Code. The interaction began with a natural language description of the desired application — CSV file upload, variable selection for dependent and independent variables, regression output with coefficients and standard errors, and a scatter plot visualization. Claude Code generated the entire application in under two minutes. When edge cases emerged (missing value handling, p-value calculations), conversational debugging resolved them in approximately five additional minutes. The total development time for a functional econometric application: under seven minutes, with zero lines of code written by the researcher.

This capability extends well beyond simple tools. The NBER paper documents that coding agents can now build data pipelines that automatically fetch and clean economic data from multiple sources, interactive dashboards for exploring high-dimensional datasets, custom statistical models that implement novel estimation procedures, and automated report generators that combine analysis with narrative interpretation. The SWE-Bench V benchmark, which measures real-world coding ability, shows GPT-5 at 75% and Claude Opus 4.1 at 72.7% — performance levels that enable genuinely complex software development through natural language alone.

Three categories of coding tools serve different researcher needs. Terminal-based agents like Claude Code, OpenAI Codex CLI, and Google Gemini CLI operate in the command line and excel at building new applications from scratch. AI-enhanced IDEs like GitHub Copilot, Cursor, and Windsurf integrate AI assistance into the existing development workflow, providing inline suggestions and multi-file editing capabilities. Cloud coding platforms like Replit and bolt.new offer browser-based environments where non-programmers can build complete web applications through conversation.

The paper raises an important equity consideration: while vibe coding democratizes access to technical capabilities, power users who combine programming expertise with agent proficiency may benefit disproportionately. As Korinek notes, “although all may benefit, those power users who know best how to deploy sophisticated AI agents may benefit far more from agentic AI than regular users.” This suggests that investing in agent literacy — even without deep programming skills — is essential for economic researchers.

Building Custom AI Agents for Economic Research

Beyond using commercial tools, the NBER paper makes a compelling case that economic researchers should build their own AI agents. This hands-on approach serves dual purposes: it develops practical understanding of agent capabilities and limitations, and it produces custom tools precisely tailored to specific research workflows.

The paper presents a clear architectural framework for AI agents: a prompt feeds into an orchestrator that coordinates a reasoning engine (the LLM) with tools (search, code execution, databases) and memory (context persistence across interactions). The ReAct framework (Reasoning and Acting) structures the agent’s operation into iterative Thought → Action → Observation cycles, where the orchestrator interrupts the LLM’s token generation to call external tools and feeds results back for continued reasoning.

The practical implementation begins with a relatively simple FRED data retrieval agent — approximately 140 lines of Python — that can answer questions like “How is the U.S. labor market doing?” by autonomously selecting relevant data series (e.g., UNRATE for unemployment), fetching recent data from the FRED API, and generating natural language interpretation. This foundational agent demonstrates the core pattern: the LLM decides which tools to call, interprets the results, and determines whether additional tool calls are needed before generating a final response.

LangGraph transforms this linear pattern into a more sophisticated state machine architecture. Three components define a LangGraph agent: state management (a centralized TypedDict that flows through the entire graph), nodes (functions that operate on state, such as think_node, act_node, observe_node, and respond_node), and conditional edges (routing logic that enables looping, branching, and dynamic behavior based on current state). The advantages over simple scripts include traceable state transitions for debugging, checkpoint and resume capability for long-running tasks, and easy extensibility through adding new nodes and edges.

Two open protocols are accelerating agent development. The Model Context Protocol (MCP), introduced by Anthropic in November 2024 and adopted by all major labs, provides a universal connection standard between AI agents and external data sources — analogized to USB-C for AI. Instead of building N×M custom integrations for N agents and M tools, MCP reduces this to N+M connections. Close to 10,000 MCP servers existed by the time of writing. The Agent2Agent (A2A) Protocol, launched by Google in April 2025 and transferred to the Linux Foundation, enables agents to communicate with each other via “Agent Cards” that describe capabilities and connection endpoints.

AI Agents Economic Research: Benchmarks That Matter

Understanding the performance benchmarks for AI agents in economic research is essential for making informed decisions about tool selection and investment. The NBER paper compiles the most comprehensive benchmark comparison available, covering language ability, reasoning depth, and coding proficiency across all major model providers.

On general language ability, measured by the LMSYS Arena (crowdsourced human preferences from millions of conversations), the leading models cluster tightly: Google Gemini 2.5 Pro (1457), OpenAI GPT-5 (1455), Anthropic Claude Opus 4.1 (1451), xAI Grok-4 (1425), Alibaba Qwen3 (1422), and Moonshot Kimi K2 (1421). The practical implication: for text-heavy tasks like summarization, editing, and drafting, switching between top-tier models produces minimal quality differences. Researchers should optimize for cost and workflow integration rather than chasing marginal quality improvements.

On reasoning and domain expertise, measured by the GPQA benchmark (graduate-level questions where PhD experts score approximately 65%), the spread is much wider: GPT-5 at 89.4%, Grok-4 at 88.9%, Gemini 2.5 Pro at 84.0%, Claude Opus 4.1 at 83.3%, and DeepSeek-R1 at 71.5%. For context, the best LLM in November 2023 scored just 39% on this benchmark. This dramatic improvement — from well below to substantially above human expert performance in approximately 18 months — underscores the acceleration in AI reasoning capabilities. For economists working on formal theory, model selection based on reasoning benchmarks is significantly more consequential than for those focused on empirical work.

On coding ability, measured by SWE-Bench V (real-world software engineering tasks), GPT-5 leads at 75%, followed by Grok-4 at 72-75%, Claude Opus 4.1 at 72.7%, Gemini 2.5 Pro at 63.8%, and DeepSeek-R1 at 49.2%. The coding benchmark is particularly relevant for AI agents in economic research because agent architecture fundamentally depends on code generation — agents must write and execute tool calls, data processing scripts, and analytical routines. A model that excels at reasoning but underperforms at coding will produce agents that think well but act poorly.

The task autonomy frontier provides perhaps the most forward-looking metric. According to research by Kwa et al. (2025), AI agents can now autonomously perform tasks that take humans about 50 minutes, and this capability horizon has been doubling every seven months since 2019. Extrapolating this trajectory, AI agents would handle day-long research tasks by end of 2026. While extrapolation carries inherent uncertainty, the consistency of this trend over seven years suggests the direction — if not the exact timeline — is reliable.

The Cost and Access Divide in AI Agent Technology

AI agents for economic research raise significant questions about equitable access to advanced research tools. The NBER paper documents a pricing landscape that spans three orders of magnitude, creating potential disparities between well-funded and resource-constrained research institutions.

At the entry level, basic AI subscriptions from OpenAI, Google, and Anthropic cost approximately $20 per month, with xAI offering Grok at $8. These tiers provide access to capable language models and limited agentic features. Premium tiers — OpenAI Pro at $200, Google AI Ultra at $250, Anthropic Max at $200, xAI SuperGrok Heavy at $300 — unlock the full spectrum of reasoning and agentic capabilities. Sam Altman has publicly discussed the possibility of selling a “PhD-level scientist system” at $20,000 per year. The tenfold price increase from basic to premium tiers, achieved in roughly one year, suggests that access to frontier AI capabilities for researchers may increasingly depend on institutional budgets.

However, powerful countervailing forces exist. Open-source models from Meta (LLaMA), DeepSeek, Alibaba (Qwen), Moonshot (Kimi-K2), and Mistral offer near-frontier performance at dramatically lower costs. The NBER paper highlights a striking example: Kimi-K2 offers API access at 15 cents per million input tokens, compared to $15 for comparable Anthropic models — a 100x price differential. For researchers willing to invest in technical setup, deploying open-source models on institutional infrastructure provides maximum security and minimal marginal cost.

The paper recommends a pragmatic approach: “maintaining flexibility across model providers rather than committing to a single ecosystem.” For most economic researchers, this means starting with a paid subscription to any leading provider for exploratory work, complementing it with API access to open-source models for high-volume tasks, and building custom agents that can switch between providers based on task requirements and budget constraints. Universities and research institutions should evaluate AI access as research infrastructure — analogous to journal subscriptions and computational clusters — rather than individual discretionary spending.

Data security considerations add another dimension to the access question. OpenAI and Anthropic provide enterprise security (SOC 2 Type 2), and API usage defaults to no training on customer content. For maximum data protection, deploying open-source models on institutional servers remains the gold standard. The National Science Foundation and other funding bodies are beginning to include AI infrastructure in grant budgets, which may help level the playing field for publicly funded research.

Limitations and Risks of AI Agents in Economics

Despite their impressive capabilities, AI agents in economic research carry substantial risks that researchers must understand and mitigate. The NBER paper is unusually candid about these limitations, devoting significant attention to failure modes that are particularly dangerous in economic contexts.

Hallucinations remain the most fundamental concern. AI systems generate plausible but incorrect content — fabricated citations, invented statistics, nonexistent datasets — while presenting them with what the paper describes as “supreme confidence.” For economic research, where empirical claims must be verifiable and citations traceable, hallucinated content can undermine entire analyses if not caught during verification. The risk is amplified in agent workflows where outputs from one stage become inputs to the next, creating potential cascades of compounding errors.

Computational cascades represent a systemic risk unique to multi-agent architectures. When a search agent retrieves incomplete data, an analysis agent may fill gaps with plausible but incorrect assumptions, and a synthesis agent may present these assumptions as established findings. Each stage compounds the error while increasing the apparent confidence of the output. The paper emphasizes that human checkpoints at each stage of agent workflows — not just at the final output — are essential for maintaining research integrity.

Brittleness and reproducibility present methodological challenges. AI agents show remarkable sensitivity to small prompt variations, producing substantially different outputs from semantically equivalent queries. This makes evaluation and reproducibility — cornerstones of scientific research — significantly harder. The American Economic Association’s data and code policies, designed for deterministic analyses, may need updating to address the stochastic nature of agent-assisted research.

Prompt injection attacks introduce security vulnerabilities into research workflows. Malicious actors can embed hidden instructions in documents or web pages that AI agents access, potentially manipulating analysis outcomes. The paper cites research by Gans (2025) showing that authors can already manipulate AI-powered peer review systems through such attacks — “a phenomenon that reveals both the speed of AI adoption in academia and the emergence of new forms of strategic behavior.”

Perhaps most concerning for economists specifically, LLMs “sometimes misapply theoretical frameworks and reproduce common misconceptions from their training data rather than maintaining rigorous economic logic.” An agent trained on internet text may confidently apply supply-demand analysis to situations requiring general equilibrium reasoning, or assume competitive markets where market power is central to the question. The paper’s recommended heuristic is powerful in its simplicity: treat AI agents “like a professor would treat a team of research assistants: they require careful planning of what is to be done, oversight during execution, and detailed vetting of the final results.”

The Future of AI Agents in Economic Research

The trajectory of AI agents in economic research points toward a fundamental transformation of the discipline’s workflow and, potentially, its intellectual structure. The NBER paper outlines several near-term developments that will accelerate adoption and capability, while raising deeper questions about the evolving role of the human economist.

On the technical frontier, the paper anticipates the development of common reusable agent patterns for core research tasks: standardized literature review agents, data analysis pipelines, and paper writing assistants that can be shared across institutions and customized for specific subfields. MCP servers from major economic data providers — FRED, the IMF, the Bureau of Labor Statistics — would enable agents to access official data through standardized protocols rather than bespoke API integrations. Cross-institutional agent collaboration, respecting data privacy constraints, could enable distributed research projects that leverage specialized agents from multiple research groups.

The emergence of markets for specialized research agents is another likely development. Just as the Stata and R ecosystems evolved package repositories where researchers share statistical methods, agent marketplaces could emerge where economists share and refine specialized tools — an agent optimized for difference-in-differences estimation, another for structural model calibration, a third for automated robustness checking.

The deeper question concerns the evolving role of the economist. The paper draws an analogy to chess: “The competition period with AI that Kasparov experienced for chess in the 1990s may be beginning for economic research.” As agents increasingly handle data collection, statistical analysis, literature synthesis, and even preliminary model specification, the economist’s comparative advantage shifts toward what remains uniquely human — ethical reasoning about social welfare, creative framing of research questions, interpretation of findings in institutional and historical context, and the exercise of judgment about what questions are worth asking in the first place.

Korinek argues that the economist’s role will evolve “from producing analysis to defining values, interpreting implications, and ensuring that economic insights serve human flourishing.” This is not a diminished role — it is arguably the most important one. The risk lies not in AI replacing economists, but in economists failing to adapt their training, methods, and professional norms to a world where the production of analysis is increasingly automated while the interpretation and application of that analysis remains irreducibly human.

For researchers and institutions navigating this transition, the NBER paper offers practical guidance: invest in agent literacy across all levels of economic training, build institutional infrastructure for secure AI deployment, maintain flexibility across providers and frameworks, and above all, develop the verification and oversight skills that transform AI agents from unreliable oracles into powerful — but properly supervised — research partners. Explore more analyses of transformative research in our interactive library.