Andreas Kostøl

AI / Research agents

Research agents

Connect your language model to research agents to learn empirical research methods and a workhorse model of wage dynamics.

Much of this site, the agents and their code was made with AI assistance. It can contain errors. Use it at your own risk, and check results before you rely on them. Feedback is much appreciated: use the Feedback button at the bottom right.

Each agent is built with Paper2Agent from the papers that describe the design. It has its own page and a set of tested tools on an MCP server. The agents for research designs run on simulated data where the true effect is known. You can then see whether a method hits the truth, and what happens when an assumption fails. The wage dynamics agent solves and simulates a published structural model.

Add the tools to an AI assistant that supports MCP with this address:

https://andreaskostol.no/mcp

The tool names begin with jd_, iv_, fe_ and bfpr_. Result files come back as download links that are deleted after a day. How to connect.

How to connect

The tools run on my server. The language model runs in your own assistant, on your own account and plan. The tools cost nothing beyond that plan and need no login or key.

  • Claude (claude.ai and the desktop app). Open Settings, then Connectors, and choose Add custom connector. Paste the address above and click Add. In a chat, turn the connector on under the + button, then Connectors. On Team and Enterprise plans an owner first adds it under Admin settings, then Connectors.
  • Claude Code. Run claude mcp add --transport http research-agents https://andreaskostol.no/mcp in a terminal.
  • ChatGPT. Turn on Developer mode under Settings, Connectors, Advanced settings. It needs a paid plan. Create a connector, paste the address and choose No authentication. In a chat, choose + and Developer mode, then switch the connector on.
  • Other MCP clients. Use the streamable HTTP transport with the address above and no authentication.

Menu names change between versions of the apps.

Then ask in plain language. Some examples:

  • "Use the jd_ tools to simulate a judge design where monotonicity fails, and compare 2SLS with UJIVE."
  • "With the iv_ tools, simulate a design with two instruments and show how far the LATE is from the ATE."
  • "Use the fe_ tools to show how few movers per firm bias the variance of firm effects, and correct it with the leave-out estimator."
  • "Use bfpr_decompose to split wage growth over the first 20 years for workers with 12-14 years of education. Then raise bargaining power with bfpr_counterfactual."

The server allows 300 tool calls an hour through this address, shared by all users, and 10,000 a day in total. If the address refuses calls, wait and try again later.

Agents

Judges, examiners and caseworkers · leniency as an instrument

Judge and examiner designs

When cases are assigned at random to judges, examiners or caseworkers who differ in leniency, leniency can serve as an instrument for the decision. The agent simulates such designs with a known true effect. It runs the checklist from the two practitioner guides with CJIVE, UJIVE and the tests for monotonicity.

Open Judge and examiner designs

Guides: Chyn, Frandsen and Leslie (2025, JEL) and Goldsmith-Pinkham, Hull and Kolesár (2026, JEP). The tools begin with jd_. Data: simulated.

Instrumental variables · when the effect varies

Instrumental variables

A valid instrument identifies the effect for the compliers, not for everyone. The agent simulates designs with a known ATE and LATE. It runs Wald, 2SLS with several instruments, the Imbens-Rubin complier distributions, marginal treatment effects and bounds, with the checklist from the Handbook chapter.

Open Instrumental variables

Papers: Imbens and Angrist (1994), Imbens and Rubin (1997), Mogstad, Torgovitsky and Walters (2021) and Mogstad and Torgovitsky (2024, Handbook). The tools begin with iv_. Data: simulated.

Firms · wage effects and sorting

Firms and wages

How much of pay comes from the firm, and how much from who works there? With few movers per firm, fixed effects overstate the variance of firm effects and understate sorting. The agent simulates linked panels with a known truth and runs AKM, FE-HO, the leave-out correction and BLM. It also runs the exercises from Bonhomme et al. and computes the markdown under monopsony.

Open Firms and wages

Papers: Kline (2025, Firm Wage Effects), Kline (2025, Handbook, monopsony), Bonhomme, Lamadon and Manresa (2019) and Bonhomme, Holzheu, Lamadon, Manresa, Mogstad and Setzler (2023). The tools begin with fe_. Data: simulated.

Wage dynamics · job search and human capital

Wage dynamics

Wages grow over a career because workers gain human capital, move to better firms and get raises when employers match outside offers. The agent solves and simulates the equilibrium search model of Bagger et al. at the published estimates or at your own values. It splits wage growth over the first 20 years into these sources. The Python code can be downloaded.

Open Wage dynamics

Paper: Bagger, Fontaine, Postel-Vinay and Robin (2014, AER). The tools begin with bfpr_. Data: the authors' published estimates and data moments, and simulations from the model.