AgentEconomist¶

external · status: active · focus: analysis · discipline: economics · started: 2026

Project page: https://arxiv.org/abs/2604.27725

Source: projects/landscape/agent-economist.yml

Positioning¶

An interactive research copilot (arXiv:2604.27725) that turns a human researcher's abstract economic intuition into a runnable agent-based computational experiment. Three stages — literature- grounded hypothesis drafting, experimental-design/simulator parameterization, and experiment execution — sit on top of AgentEconomy, an LLM-driven agent-based macroeconomic simulator (households, firms, government, banking, calibrated from empirical microdata). A human reviews and steers each stage.

Distinctive contribution¶

Targets the ideation-to-simulation slice of the economics research pipeline specifically, pairing a 13,000+ paper domain knowledge base with a full agent-based macro simulator rather than a generic code sandbox; explicitly positions itself against "The AI Scientist"-style systems, arguing those optimize for an autonomous end product over supporting the human's own research process.

Evaluation scores¶

Dimension	Score (0–3)	Note
Lifecycle coverage	2	Four stages — hypothesis drafting, experimental design, simulator parameterization, and execution/results — but stops before paper-drafting or referee-simulation.
Autonomy level	1	Explicitly interactive/human-in-the-loop: the researcher reviews and steers each of the three pipeline stages rather than approving only a final artifact.
Architectural transparency	1	Paper describes the three-stage architecture and the AgentEconomy simulator design; no code, prompts, or configs found public at scoring date.
Inputs supported	1	Single input form (a free-text economic intuition), but draws on both a literature corpus (13,000+ papers) and empirical microdata for simulator calibration.
Outputs / reproducibility	2	Persists structured experiment artifacts (simulator configs, run code, results) per the paper, but the system itself is not public, so end-to-end reproduction by third parties is unverified.
Internal evaluation	2	Systematic evaluation comparing AgentEconomist-produced ideas/experiments against baseline-LLM outputs using human expert judges on the same prompts; arXiv preprint only, not yet peer-reviewed at scoring date.
Openness	0	No public code or AgentEconomy simulator release found at scoring date; paper only.
Maturity / traction	1	Academic research prototype (arXiv, posted 2026-04-30); single-team, no reported external users or releases.
Cross-family policy	0	No cross-model-family review described; single underlying LLM backend.
Runtime assurance	1	Literature-grounding of hypotheses against the domain KB acts as a light in-pipeline check; no described math/claim-audit or adversarial gate.
Cross-platform portability	0	Research prototype tied to the authors' AgentEconomy simulator; no described deployment on other runtimes.

Scored on 2026-11-01. See the evaluation rubric.

Tags¶

Pipeline stages: hypothesis-generation research-design formal-modeling data-analysis

Architectural features: multi-agent human-in-loop rag-knowledge-base tool-use iterative-loop

Inputs: research-direction

Outputs: computational-results simulation-code

Data sources: empirical-microdata

Knowledge sources: economics-literature-corpus

Limitations¶

No public code or data release found at scoring date — architecture and results are known only from the paper.
Scope is agent-based macroeconomic simulation specifically, not general empirical economics (no identification strategy, no real-dataset estimation, no paper-drafting stage).
Evaluation relies on a human expert user study of unreported scale; quantitative headline results were not found in public summaries at scoring date.

Papers describing this project¶

AgentEconomist: An End-to-end Agentic System Translating Economic Intuitions into Executable Computational Experiments — Chen, Jiaju, He, Xiangnan, Li, Yong, et al. (2026). arXiv. arXiv:2604.27725

chen2026agenteconomist (BibTeX)