AI Agents Transforming Economic Research: NBER Working Paper 34202 Analysis

Three AI Paradigms Economists Must Know

The landscape of artificial intelligence has evolved from simple chatbots to sophisticated autonomous agents, fundamentally altering how economic research can be conducted. According to NBER Working Paper 34202 by Anton Korinek, we’re witnessing three distinct paradigms that economists must understand to leverage AI effectively.

Traditional large language models (LLMs) like early ChatGPT represent the first paradigm—fast, intuitive responses that mirror human System 1 thinking. These models excel at quick analysis and writing tasks but lack the deliberate reasoning needed for complex economic problems.

Reasoning models constitute the second paradigm, embodying System 2 thinking through deliberate, step-by-step analysis. These models can now achieve gold-medal performance at the International Mathematical Olympiad and score 89.4% on graduate-level reasoning benchmarks—significantly outperforming the ~65% achieved by PhD-level human experts.

The third paradigm—agentic AI—combines both systems into autonomous agents that can plan, execute, and adapt their approach to achieve research goals. These agents can break down complex tasks, use multiple tools simultaneously, and iterate on their strategies without constant human guidance. For economists, this represents a shift from using AI as a sophisticated autocomplete tool to employing it as a research collaborator capable of independent analysis.

Deep Research Agents Synthesizing Hundreds of Sources

The most dramatic transformation in academic workflow comes from Deep Research agents that can autonomously conduct comprehensive literature reviews. These systems access and evaluate more than 500 internet sources per assignment, producing detailed reports in 5-30 minutes that would previously require weeks of manual work.

The cost-effectiveness is remarkable: a self-built Deep Research agent costs approximately 1 cent in OpenAI tokens plus 15 Tavily search queries per report. This represents a fundamental shift in the economics of research preparation, making comprehensive background analysis accessible for any research question, regardless of budget constraints.

What makes these agents particularly valuable for economists is their ability to synthesize disparate sources and identify patterns across large bodies of literature. They can trace theoretical developments, identify contradictory findings, and highlight emerging research directions—tasks that traditionally required extensive manual reading and note-taking. However, the paper emphasizes that human oversight remains crucial for evaluating the quality of sources and ensuring the economic reasoning is sound.

Vibe Coding: Building Tools Without Programming

Perhaps the most accessible advancement for economists without programming backgrounds is “vibe coding”—using AI to build functional research tools through natural language descriptions rather than formal coding knowledge. The paper demonstrates this concept with working implementations of research agents, including a FRED data retrieval tool built with approximately 140 lines of largely AI-generated Python code.

This approach democratizes the creation of custom research tools. Economists can now describe their data needs, analytical requirements, or workflow preferences, and AI agents can generate working code that implements these specifications. The paper shows examples ranging from simple data extraction scripts to sophisticated multi-agent research systems that coordinate multiple AI models to accomplish complex analytical tasks.

The implications extend beyond individual productivity. Research teams can rapidly prototype analytical tools, test different methodological approaches, and customize data processing workflows without requiring dedicated programming resources. This lower barrier to entry for research automation could accelerate the pace of economic analysis and enable more experimental approaches to methodology development.

AI Lab Benchmark Scorecard and Rising Costs

The competitive landscape among AI providers has intensified dramatically, with six major labs now competing at the frontier: Google DeepMind, OpenAI, Anthropic, xAI, Alibaba/Qwen, and Moonshot AI. The paper provides detailed benchmark comparisons that economists should understand when selecting research tools.

On the LMSYS leaderboard, top traditional LLMs cluster tightly between 1421 and 1457 points, suggesting that high-quality language models have essentially become commodities. The real differentiation now occurs in reasoning capabilities and specialized tasks. GPT-5 leads the GPQA (graduate-level reasoning) benchmark at 89.4%, followed closely by xAI’s Grok-4 at 88.9%.

However, the democratization of AI capabilities faces a significant challenge: rising costs. Premium AI subscriptions have increased tenfold from $20 standard subscriptions to $200-$300 monthly for advanced tiers. A researcher subscribing to the most expensive tier at each leading lab would spend close to $1,000 per month, with OpenAI reportedly considering a $20,000 annual “PhD-level scientist” subscription.

This cost escalation raises important concerns about access inequality in academic research. Well-funded institutions and researchers may gain significant advantages in research productivity, while budget-constrained researchers could be excluded from frontier AI capabilities.

Open-Source Models Closing the Gap

Open-source AI models are emerging as powerful alternatives to expensive proprietary systems. Models from DeepSeek, Alibaba’s Qwen, and Kimi-K2 offer competitive performance at dramatically lower costs. Kimi-K2, for example, provides comparable capabilities at 15 cents per million input tokens versus Claude’s $15—a 100× price differential.

This cost advantage comes with trade-offs. Open-source models may require more technical setup, lack some of the polish and safety features of commercial systems, and might have slightly lower performance on cutting-edge tasks. However, for many research applications—data processing, literature analysis, preliminary coding—the performance gap is often negligible while the cost savings are substantial.

The paper emphasizes that open-source deployment offers the gold standard for data security, allowing researchers to process sensitive information without sending it to external services. For economists working with confidential data or conducting research for government or corporate clients, local deployment of open-source models may be the only viable option for AI-assisted research.

Building Research Agents Step-by-Step

The paper provides concrete implementation guidance through working code examples, including a comprehensive multi-agent Deep Research system built with approximately 370 lines of Python. The approach uses the ReAct (Reasoning and Acting) framework combined with LangGraph’s state-machine architecture to create agents that can plan, execute, and adapt their research strategies.

The implementation process involves defining agent roles, establishing communication protocols between agents, and creating feedback loops that allow the system to refine its approach based on intermediate results. For economists, this represents a practical pathway from concept to working research tool, with detailed examples showing how agents can coordinate to accomplish complex analytical tasks.

Protocols Connecting AI to Economic Data

The infrastructure supporting AI research agents has expanded rapidly, with close to 10,000 Model Context Protocol (MCP) servers available less than one year after the protocol was established. For economists, this means AI agents can now directly connect to a vast ecosystem of data sources, analytical tools, and research platforms.

The Agent-to-Agent (A2A) protocols enable AI systems to coordinate complex research workflows by dividing tasks among specialized agents. One agent might focus on data collection from Federal Reserve databases, another on statistical analysis, and a third on literature synthesis. This division of labor mirrors how human research teams operate but with the speed and consistency of automated systems.

Where AI Agents Fall Short and Security Considerations

Despite impressive capabilities, AI agents face significant limitations that economists must understand to use them effectively. The paper identifies several critical areas where human oversight remains essential, starting with the persistent problem of hallucinations—AI agents generating plausible but incorrect content with high confidence.

Computational cascades represent another serious risk, where errors propagate through multi-agent workflows without detection. When one agent’s incorrect output becomes another agent’s input, errors can compound rapidly, leading to conclusions that appear well-reasoned but rest on fundamentally flawed premises.

The security implications of AI-assisted research require careful consideration, particularly for economists working with sensitive data. Prompt injection attacks pose a growing concern as AI agents become more sophisticated and autonomous. For maximum data security, the paper recommends deploying open-source models on internal infrastructure, though this requires substantial technical resources.

The Automation Horizon and Future Implications

The trajectory of AI capabilities suggests dramatic changes in how economic research will be conducted. With task competence doubling every seven months since 2019, AI agents could autonomously perform day-long research tasks by the end of 2026. This exponential improvement curve indicates that the current limitations of AI research tools may be temporary constraints rather than permanent boundaries.

The paper projects that by June 2025—just nine months after the first public reasoning model—advanced AI systems could solve some of the hardest solvable mathematical problems in existence. For economists, this suggests that complex analytical tasks requiring sophisticated quantitative reasoning may soon be within reach of AI agents, fundamentally changing the nature of economic research work.

This automation horizon raises important questions about the future role of economists in research production. As AI agents become capable of conducting increasingly sophisticated analysis independently, the value of human economists will likely shift toward defining research questions, interpreting results in broader contexts, and ensuring that research serves human welfare goals.

Practical Recommendations for Adoption

For economists ready to integrate AI agents into their research workflows, the paper provides specific actionable recommendations. Start with a paid subscription to any leading AI provider—maintaining flexibility across providers is advised over committing to a single ecosystem. The modest cost of basic subscriptions represents a reasonable investment in understanding AI capabilities and limitations.

Begin with low-stakes experimentation where verification is straightforward. Use AI agents for preliminary literature searches, data collection tasks, and initial analytical exploration. As confidence and understanding grow, gradually expand to more complex workflows while maintaining robust verification processes for all AI-generated content.

Institutions should base AI compliance policies on evidence-based security assessments rather than outdated assumptions about AI capabilities. The paper emphasizes that blanket bans on AI tools often do more harm than good, preventing researchers from developing crucial skills while failing to address legitimate security concerns.

Most importantly, economists should view this technological shift as an opportunity rather than a threat. AI agents can handle routine analytical tasks, freeing economists to focus on higher-level questions about research design, theoretical innovation, and policy implications. The goal is not to replace human insight with artificial intelligence, but to augment human capabilities with powerful tools that accelerate the pace of economic discovery and analysis.