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

TEXT START: 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.

The Dissection

The paper is designing an auditable AI production line for economic research. It decomposes ideation, literature review, modeling, data work, implementation, estimation, and reporting into seven specialized agent teams coordinated by an orchestrator, with human review at each stage.

Its real contribution is operational, not epistemic: it attempts to make AI-generated research traceable, bounded, reproducible, and institutionally acceptable. Provenance logs, authority limits, execution checks, causal-identification requirements, simulations, counterexample searches, and LLM review ensembles are control mechanisms for deploying synthetic researchers at scale.

The phrase “human-in-the-loop” is the load-bearing legal and reputational brace. It signals that humans still certify the system, even as the workflow is built to compress the amount of human labor required per research project.

The Core Fallacy

The paper risks confusing verification with human economic indispensability.

A workflow can verify provenance, test code, inspect identification strategies, simulate models, and search for counterexamples without preserving the economist’s position in the production circuit. It can make research safer while making research labor less necessary. Under the Discontinuity Thesis, that is not a contradiction. It is the mechanism.

The methodology addresses whether an AI system can produce acceptable research artifacts. It does not address who continues to earn wages when the system produces those artifacts faster, cheaper, and with fewer researchers. Quality control is not productive participation. Human approval is a lag defense, not a permanent moat.

The central structural outcome is therefore inverted from the paper’s implicit institutional framing: the more successfully the workflow records and bounds agent behavior, the easier it becomes to standardize, transfer, audit, and eventually automate the remaining human oversight.

Hidden Assumptions

• Human reviewers will remain affordable and necessary at every stage rather than becoming a final-signoff bottleneck targeted for automation.
• Human experts can reliably detect errors in AI-generated reasoning, even when the workflow’s complexity exceeds any individual reviewer’s full understanding.
• Provenance and verification establish trustworthiness rather than merely documenting a chain of potentially coordinated mistakes.
• A dual-judge LLM ensemble can meaningfully evaluate agent outputs without reproducing the same model failures in a more ceremonial form.
• Explicit checks for identification, stylized facts, and counterexamples capture the full substance of economic reasoning.
• A prototype produced in hours constitutes research progress rather than merely accelerating the production of plausible, reviewable artifacts.
• Open-weight local models solve the strategic problem. They reduce dependence on external providers, but they do not eliminate compute, data, maintenance, coordination, or capability gaps.
• Dividing work among specialized teams reduces complexity rather than creating new interfaces where errors propagate and accountability fragments.
• The discipline’s current standards are stable enough to serve as the control specification for systems that may eventually reshape the discipline itself.
• Faster research output will expand opportunity for economists rather than increase supply, intensify competition, and collapse the wage value of routine research labor.

Social Function

Primary classification: transition management and partial truth.

The paper correctly recognizes that general-purpose LLMs are unreliable without discipline-specific controls. It offers institutions a way to absorb the technology without immediately surrendering standards, accountability, or legitimacy.

But it also functions as elite self-exoneration. By emphasizing human review, provenance, and trustworthy execution, it allows institutions to present automation as responsible augmentation while quietly constructing the infrastructure for labor substitution. The human remains visibly present as a guarantor while becoming less central to the actual production process.

Its prestige function is equally clear: the workflow converts raw automation into the language of methodological rigor. That is useful for adoption. It does not alter the underlying ownership question. Whoever controls the models, compute, data, orchestration layer, and institutional access captures the surplus. The researchers operating inside the workflow become Servitors unless they control those assets or become indispensable to them.

The Verdict

This is not a defense of the economist. It is a blueprint for an auditable economic-research factory.

Its safeguards may delay institutional panic and reduce obvious error, but they do not defeat cognitive automation, coordination pressure, or the collapse of productive participation. The paper’s success condition—reliable research prototypes produced in hours—is precisely the evidence that the old research labor circuit is being severed.

Under the Discontinuity Thesis, the workflow is transition infrastructure: valuable to its owners, temporarily useful to specialized overseers, and ultimately hostile to the mass of researchers whose work it systematizes.