Introduction
This is the final part of a five-part series on applying AI agents to empirical research. The core argument: empirical research has five bottlenecks—finding data, running regressions, writing papers, responding to reviewers, and defending the work. Each used to be manual drudgery. MCP servers shrink data acquisition from a month to one command, a Stata Skill replicates an AER paper in 10 minutes, and multi-agent systems compress reviewer responses from 6–8 months to 2 weeks. AI doesn't replace researchers; it returns drudgery time so researchers keep the core judgment.
Why empirical research is slow
The traditional pipeline: topic selection → literature search → data acquisition → cleaning → regressions → robustness checks → drafting → revision → submission → reviewer response → defense. Typical costs: data acquisition takes weeks to months; regressions days; replication weeks; reviewer responses 6–8 months.
The core problem: AI knows "how" but not "the full workflow." Ask it for a DID analysis and it stops at the baseline regression, then grudgingly adds parallel-trends tests and placebo checks one at a time. Skills encode senior researchers' methodology into structured workflows—teaching AI to *do research*, not just write code.
Data acquisition: from a month to one MCP command
| MCP server | Scale | Coverage | |---|---|---| | FRED | 800k+ economic time series | Federal Reserve, US macro | | World Bank | 1,000+ indicators, 200+ countries | Global development | | US Gov Open Data | 40+ APIs | Census/CDC/FDA/FEC etc. | | French DataGouv | National open-data platform | French official statistics |
Chinese microdata panels—CFPS, CGSS, CHARLS, CHIP, CHFS—are also accessible via MCP interfaces or aggregators. OpenAlex aggregates 240M scholarly works; combined with Semantic Scholar, arXiv, PubMed, and CrossRef, it forms a scholarly data federation. Renmin University's DeepAnalyze (4,000+ stars) builds on this: given a research question, it retrieves literature, locates data, runs analysis, and writes reports.
Replication: an AER paper in 10 minutes
A Stata Skill parses .do files line by line and maps commands to Python equivalents (reghdfe → pyfixest, ivreg2 → linearmodels), automatically inferring panel identifiers, time variables, and clustering levels, then aligning outputs (coefficient tables, robustness checks, heterogeneity analysis).
Three replication paths:
- Stata → Python: Stata Skill + translation layer (e.g., reviewer requests)
- Explicit Python: Full Empirical Analysis Skill (teaching, auditing)
- Python DSL: StatsPAI (900+ functions, "agent-native" design)
- ClawHub: 13,729 Skills, one-command install
- VoltAgent: 5,400+ curated Skills
- antigravity: 1,340+ Skills (28,000 stars),
npx antigravity-awesome-skills - Main repo: https://github.com/brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research
- CoPaper.AI: https://copaper.ai
- Agent Laboratory: https://github.com/SamuelSchmidgall/AgentLaboratory
- AI-Scientist-v2: https://github.com/SakanaAI/AI-Scientist-v2
- DeepAnalyze: https://github.com/ruc-datalab/DeepAnalyze
- HKU AI-Researcher: https://github.com/HKUDS/AI-Researcher
- ClawHub: https://clawhub.com
- FRED MCP: https://github.com/stefanoamorelli/fred-mcp-server
- OpenAlex MCP: https://github.com/oksure/openalex-research-mcp
- Security scan report: https://github.com/brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research/blob/main/SECURITY-SCAN-REPORT.md
Three autonomous research systems:
| System | Stars | Key result | |---|---|---| | Agent Laboratory | 5,500 | Research cost reduced 84% | | AI-Scientist-v2 | – | First AI-written paper to pass peer review | | ARIS | – | Overnight autonomous research |
Reviewer response: from 6–8 months to 2 weeks
The review-response workflow: feed the reviewer PDF to AI → classify each comment (major/minor/clarification/suggestion) → draft replies → flag required additional analyses with generated code. Supporting tools include AI-research-feedback (6 AI agents simulating top-journal reviewers across identification credibility, causal overclaiming, robustness, contribution clarity, methodological rigor, writing quality) and paper-slide-deck (one-click paper-to-PPT conversion).
Multi-agent systems: role separation = peer-review logic
A single agent that writes and reviews its own work tends to approve itself. Role separation creates a genuine quality loop. CoPaper.AI uses a Supervisor plus four sub-agents—Preparation (cleaning, variables), Modeling (20 causal methods: DID/IV/RD/PSM/DML), Visualization (publication-grade tables), Writing (citations, academic prose)—with intelligent task routing. HKU's AI-Researcher (NeurIPS 2025 Spotlight) runs literature review → hypothesis generation → algorithm implementation → LaTeX paper writing autonomously. Renmin University's DeepAnalyze supports a full-Chinese environment (Chinese variable names, comments, reports).
Aggregation platforms
Supporting infrastructure: Anthropic's 32-page Claude Code Skills guide, CEPR/VoxDev and HKU lectures, causal-inference textbooks, and a security audit (52/52 local Skills CLEAN).
Critical boundaries
1. Data permissions: CFPS/CHARLS require account applications AI can't complete; commercial databases (Wind, CEIC, Bloomberg) need institutional subscriptions; APIs have rate limits. 2. Equivalence trap: Stata→Python translation ensures *numerical* equivalence but not always *methodological* equivalence—convergence criteria, small-sample corrections, and multi-way clustering differ across packages. Automated output needs human verification, especially when results "look right." 3. Coordination costs: multi-agent systems face communication overhead, error propagation (bad cleaning corrupts modeling), and debugging complexity. Fast paper-generation likely reflects happy-path data. 4. AI-Scientist-v2's milestone: which journal? Did reviewers know it was AI-written? Passing desk screening at a low-barrier outlet differs from acceptance at a top journal—its publicity value may exceed academic value. 5. Ecosystem fragmentation: 20,000+ Skills exist, but most are never reviewed; some are stale prompt wrappers or security risks.
Conclusion
A researcher's day in 2026: pose a question at 09:15; AI pulls FRED/CPS data, runs the full DID suite, generates publication-grade tables by 09:45; the human reviews identification strategy and results; six simulated reviewers pre-screen before submission; reviewer responses are drafted and executed within hours. This is not science fiction—most components already exist.
The key point: AI compresses drudgery from months to hours, but "thinking time"—whether the question matters, whether the identification is credible, what the policy implications are—remains human territory, and becomes more precious because you finally have time to think.