AiEconLab (AEL)
A permanent virtual research team for applied economists. Eight core agents, twelve experts, and a plan-time consultant designed from first principles for econ review.
View on GitHub → AiPlus platform
Apache-2.0 v0.1.0 Python + Stata + LaTeX Built on AiPlus
Why this exists
A single AI coding agent, in a long research session, drifts between roles. The prompt history mixes identification reasoning with Stata syntax with rebuttal language; each task gets done shallowly. AiEconLab replaces that single agent with a small standing team where each member has its own persona, workspace, and memory.
You talk only to Advisor (strategic) and PI (execution). The PI dispatches the rest: Theorist, PM, RA-Stata, RA-Python, Referee, Replicator. Twelve experts sit dormant until summoned. STOP-gates (journal submission, R&R response, data sharing, authorship changes) always escalate to you.
What's distinct about AEL
LLM-as-Measurement Specialist
For papers using LLMs to score text data — sentiment, ideology, classification. Designs the multi-model cross-validation panel, hand-coded subsample, held-out test docs, prompt versioning, leakage prevention — before the first scoring run.
Research-tuned consultant
Five plan-time seats: Design Credibility · Contribution Framing · Day-1 Reproducibility · IRB / Disclosure Gate · LLM-as-Measurement. Replaces AiPlus's default SWE consultant when AEL is installed.
Git-worktree workspaces
Each code-touching role gets its own worktree on its own branch. RA-Stata and RA-Python work in parallel; conflicts surface through git, not silent context overwrites.
Local-first, audit-able
All agent state, personas, memories, and dispatches live in
your project's .aiplus/. No cloud sync. No
telemetry. No upload. Records are plain JSONL — greppable,
git-able, replicable.
Install
Requires AiPlus ≥ 0.5.5.
# macOS Apple Silicon
curl -L https://github.com/izhiwen/AiPlus/releases/latest/download/aiplus-aarch64-apple-darwin.tar.gz | tar xz
sudo mv aiplus /usr/local/bin/
# In your research project
cd MyResearchProject
aiplus add aieconlab
aiplus install codex # or claude-code, opencodeOther platforms (Intel Mac, Linux x86_64, Linux aarch64, Windows): grab a binary.
Worked example
Multi-LLM-Validation-Demo — 294 nineteenth-century Classical Chinese archival documents scored across five frontier LLMs (GPT, Claude, Gemini, Qwen, DeepSeek); pairwise correlations 0.85–0.95. The validity backbone for the maintainer's job market paper, and the executable instance of the LLM-as-Measurement Specialist's protocol.
Honest disclosure
- v0.1.0. Structural acceptance tests pass; install flow works on five platforms; CI green.
- A complete real-world cycle from paper kickoff through submission has not yet been completed using AEL. Design vs reality gap is unknown.
- Single maintainer (Steve Zhiwen Wang, applied-econ PhD candidate at Pitt).
- Three v0.2 experts (Survey/Experiment, Computation, Co-Author Liaison) ship as config stubs in v0.1.
More
- DESIGN.md — full design rationale, role mapping, routing protocol
- CHANGELOG.md
- SECURITY.md — data boundaries, STOP-gates, IRB policy
- CONTRIBUTING.md