“Not Future Possibilities but Present Realities”: How AI Agents Are Reshaping Economic Research

Economics in the Age of AI Agents

Over the past two years, the way economists conduct research has undergone a profound structural transformation. In his 2025 NBER Working Paper AI Agents for Economic Research, Anton Korinek argues that the research community is moving from simple conversational systems such as ChatGPT to a new paradigm centered on autonomous AI agents. These systems integrate text generation with planning, memory, and tool use, enabling them to execute multi-step research tasks. More broadly, the rise of AI agents represents a technological reconstruction of how research itself is organized.

Korinek reflects on his earlier paper Generative AI for Economic Research (Korinek, 2023) and notes that within just two years, AI tools have evolved from conceptual prototypes into mature systems ready for daily research workflows. They no longer merely compose emails or summaries but independently conduct literature reviews, debug econometric code, run models, and interpret results. He describes this as ‘fundamental architectural advances in how AI systems operate’, where speed, cost, and organization are being re-defined by AI systems.

I. The Three Paradigm Shifts of AI: From Language to Reasoning to Autonomous Action

Korinek divides the evolution of AI into three paradigms and argues that this trajectory mirrors the methodological progress of economics itself, from automated language handling to verifiable reasoning and finally to agents capable of autonomous task execution.

The first paradigm comprises traditional large language models (LLMs) such as GPT-3 and ChatGPT. These models excel at producing coherent text and responding to natural-language queries but remain essentially single-turn reactive systems. Korinek invokes Kahneman (2011)’s concept of “System 1” thinking to liken them to human intuition — fast and pattern-recognizing yet lacking self-reflection and logical verification. Their errors in mathematical reasoning stem from the absence of an explicit reasoning chain.

The second paradigm is that of reasoning models. Since September 2024, new models have incorporated Chain-of-Thought and step-by-step verification mechanisms, allowing AI to perform Kahneman’s “System 2” reasoning. For economics, this is profound: AI can now solve intertemporal optimization, carry out dynamic programming, and generate operational econometric code. Korinek emphasizes that enhanced reasoning capability elevates AI from linguistic assistance to conceptual participation in research design and theoretical verification.

The third paradigm, which is called agentic chatbots, emerged at the end of 2024. These models do not merely understand and reason but can also act. They plan tasks, invoke external tools, search the web, execute Python scripts, and adjust their strategies based on intermediate feedback. Korinek argues that this marks a shift from AI as a passive responder to AI as an active collaborator in research, capable of handling entire workflows from data retrieval to model execution. For economists, this represents a fundamental change in the division of research labor.

From Reasoning Models to Autonomous Agents

II. Breakthrough Applications and the Construction of AI Agents for Economic Research

By 2025, the application of AI agents in research had moved beyond experimentation. Korinek highlights two representative systems: Deep Research and Claude Code, which demonstrate breakthroughs in knowledge integration and automated programming respectively.

The Deep Research system, introduced by Google DeepMind in late 2024, embodies the idea of allowing AI to perform multi-stage searches, filtering, synthesis, and structured report generation across web and database sources. Korinek notes that such systems can retrieve hundreds of papers and produce citation-rich reports within minutes. While dramatically improving the speed of information processing, two limitations remain: first, reliance on existing knowledge restricts originality; and second, quality discrimination across frontier and peripheral literature is still imperfect. Korinek concludes that Deep Research represents the first step toward automating information synthesis in economics rather than a replacement for scholarly judgment.

Another major innovation comes from Anthropic’s Claude Code, centered on the idea of “vibe coding” which describes the desired functionality in natural language so that the AI automatically generates, debugs, and optimizes complete code modules. Korinek calls this a “linguistic turn in programming,” with an impact comparable to the introduction of early statistical software. In economic applications, Claude Code can generate regression scripts, structural modeling routines, or simulation programs through a few rounds of dialogue. This substantially lowers entry barriers: economists without coding expertise can now perform complex computational analyses. Korinek suggests that the skill composition of economics is shifting, as programming proficiency gives way to prompt engineering and conceptual modeling.

In Section 4, Korinek demonstrates how to construct customizable research agents. Using the LangGraph framework, researchers can generate and connect task nodes through natural language descriptions. For example, “download data from FRED,” “run regression,” “write results summary.” These agents autonomously determine subsequent actions based on intermediate outputs, exhibiting genuine autonomy. Korinek terms this phenomenon an “AI self-replication loop”: AI systems are now assisting humans in writing the next generation of AI. As standard protocols such as MCP (Model Context Protocol) and A2A (Agent-to-Agent) evolve, collaboration across agents becomes seamless, signaling the emergence of a multi-agent ecosystem for economic research.

III. Technological Democratization and the Return of Cost Inequality

Korinek emphasizes the dual effect of AI democratization: while it lowers entry barriers, it also amplifies inequality in effective use. Researchers skilled in prompt design and model structure achieve higher productivity, leading to divergence in research speed and quality, a phenomenon which is called “inequality behind democratization.”

He further analyzes the economic constraints. As computational demands increase, premium subscriptions for closed-source frontier models such as GPT-4 or Claude 3.5 have risen to $200 per month that has become ten times higher than in 2024. Access to intelligence thus becomes a function of ability to pay. At the other end, open-source models such as DeepSeek, Alibaba’s Qwen, and Moonshot’s Kimi-K2 deliver near-frontier performance at minimal cost, acting as an equalizing force. The author advises the economics profession to maintain cross-platform flexibility and avoid dependence on single ecosystems to prevent new forms of technological monopoly.

Deep Research and Vibe Coding Building the Future of Economic Inquiry

IV. The Limitations of AI Agents and the Need for Human Oversight

Despite their promise, Korinek repeatedly cautions researchers. First, AI agents still suffer from “hallucinations”: they may fabricate citations or data. When operating in multi-agent environments, such errors can cascade. Second, they are highly sensitive to prompts, lacking reproducibility, and remain vulnerable to prompt-injection attacks. Moreover, they sometimes misapply theoretical frameworks or replicate biases from training data.

Therefore, it is suggested that economists treat AI agents as they would human research assistants: set clear objectives, supervise execution, and verify results. In other words, AI should be treated as a set of research collaborators to be managed and evaluated, not as unquestionable authorities. Such supervision safeguards research integrity and prevents excessive automation.

V. Conclusion: The Structural Transformation of Economic Research through AI Agents

In the end, Korinek offers a measured outlook for future research. The real potential of AI agents lies not merely in productivity gains but in transforming the organizational logic of research. They shift research from linear workflows to networked collaboration, from individual labor to human-machine hybrid teams. As agents acquire the ability to plan, execute, and learn, the boundaries of scholarly inquiry are being redrawn.

He concludes: “Economists should employ AI agents as professors do their research assistants, along with clear planning, supervision, and careful vetting of results.” AI may not yet produce independent theoretical innovations, but it is already reshaping the production function of research itself. The economists of the future must not only understand models but also master collaboration with intelligent agents.

References:

Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux.

Korinek, A. (2023). Generative AI for Economic Research: Use Cases and Implications for Economists. Journal of Economic Literature, 61(4), 1281–1317. https://doi.org/10.1257/jel.20231736

Korinek, A. (2025). AI Agents for Economic Research. NBER Working Paper Series. https://doi.org/10.3386/w34202