Agentic Workflows for Economic Research: Design, Implementation, and Evaluation

Abstract

Large Language Models (LLMs) can now assist across the economic research lifecycle, yet general-purpose LLMs and agent systems give no assurance that their outputs meet the discipline's bar. We present an agentic-workflow methodology that carries a research question through the full lifecycle, from ideation and literature review through modeling, data, implementation, estimation, and reporting, across seven specialized teams under a cross-team orchestrator, with human-in-the-loop review at each stage. The methodology incorporates two commitments. The first, trustworthy execution, is a shared infrastructure that verifies outputs, records the provenance of every step, bounds each agent's authority, addresses agent-specific threats, and serves every LLM locally on open weights. The second, reasoning as an economist, makes disciplinary checks explicit in the workflow: agents state identification strategies for causal claims, check structural models against stylized facts by simulation, and search for counterexamples to theoretical claims. Together they produce a first research prototype in hours, with every step's provenance and verification verdict recorded. A two-tier framework evaluates the workflows across fourteen dimensions, combining execution-log metrics with a dual-judge LLM-as-reviewer ensemble across all seven teams.

arXiv commentVersion 2 extends the v1 design to an implemented, open-source workflow (Agentic Econ Lab: seven teams under a cross-team orchestrator) and adds a two-tier evaluation over 26 chained runs. 53 pages, 9 figures, 11 tables. Code: https://github.com/AgenticEconLab/agentic-econ-lab ; run outputs and expert ratings: https://doi.org/10.5281/zenodo.23258640