Home » Departments and Centres » Academic catalogue » Course detail

Open sidebar menu

Agentic AI for Economic Research (ECO-AD-AGENTICAI)

ECO-AD-AGENTICAI

Purpose

This course provides practical training in the use of agentic AI for economic research. The main goal is to help researchers use AI to undertake more ambitious empirical and theoretical projects, with greater rigour and reproducibility, while maintaining their understanding of the work. The course is hands-on and centred on research workflows with Codex, Claude Code and similar agentic tools.
This outline is preliminary. The topics below indicate the intended coverage rather than a fixed sequence of sessions. Their weight, the specific tools and the examples will be adjusted after a pre-course survey of participants’ current use of AI and the practices they would like to develop.

Indicative topics:
Models and agentic environments
We will begin with an overview of current frontier models, their relative strengths and the trade-offs between cost and performance for researchers with limited budgets. Much of what makes these models useful for research comes from the harness around them. We will show how Codex and Claude Code let a model edit project files, run code and check its output, and how to coordinate several agents working in parallel. We will also cover the desktop apps, which are developing rapidly and make it easier to run sessions side by side, review agents’ changes and view results such as HTML reports without leaving the working environment.

Delegation and the researcher’s role
Working with agents can place researchers in a role closer to that of a principal investigator, directing work carried out by others. This makes decisions about what to delegate and how closely to supervise it more important. We will discuss how these decisions depend on the researcher’s expertise and on how readily the work can be checked. Examples will illustrate how to give agents room to work while keeping control of research decisions and a clear understanding of what has been done.

Research organisation with Git and GitHub
AI makes it easier to build well-organised, reproducible research projects from the outset. It also makes that organisation more important, because agents generate code and output quickly and can leave a project cluttered with redundant scripts, intermediate files and superseded versions. We will discuss how to structure project folders, with a focus on using Git and GitHub to back up work and safely combine contributions developed in parallel. Practical examples will cover using issues to define and track agent tasks and pull requests to review their changes before merging, with worktrees allowing several agents to work on the same project at once. We will also show how to maintain a useful record of what changed and why, so that researchers and agents can understand earlier decisions and build on them.

Context and reusable research workflows
Agents are powerful partly because they can draw on a large amount of context about the user and the project when tackling a task. Providing the right information in the best format matters both for their performance and for their cost. We will discuss how best to do this, starting with AGENTS.md files for project instructions and user preferences.

We will also discuss how to create skills that automate recurring tasks as well as more complex research workflows, from brainstorming and planning through to review. Skills can also impose taste and set standards for the code agents write, which is otherwise often unreadable and hard to maintain. They can ask agents to test what matters for economists rather than the tests written in software engineering, on which these models are largely trained. For example, in empirical work, agents can check that merges neither drop nor duplicate observations, that each firm or household appears once per period, that weighted totals match published aggregates and that shares lie between zero and one. In quantitative models, they can check that the solution has converged, that Euler equation residuals are small, that markets clear and that the model reproduces known special cases. Plugins and subagents will be introduced where they help carry out these workflows.

Supervising and verifying agent work
We will discuss how to supervise agents and verify their work, including asking them to prepare HTML reports on intermediate results so that researchers can inspect the analysis as it develops. To counter the tendency of models to agree with the user, separate agents can be given an adversarial role, checking outputs and challenging the proposed approach. Examples will show how to direct these reviews and resolve disagreements between agents while recognising where the researcher’s own judgement is needed.

Working with very large datasets
Agents make tools such as DuckDB and Polars more accessible to researchers who have not previously used them. We will show how to have agents write DuckDB SQL or Polars lazy queries, in which the full sequence of operations is optimised before execution so that only the necessary data are read and processed. Practical examples will show how these tools reduce running time and memory use, and how to verify that the faster code reproduces the original results.

Working with agents in restricted-data environments
Researchers often work with confidential data held on secure servers or in environments without internet access. We will discuss how to use agentic tools in this setting, even when the agents cannot access the data themselves. One example is using agents to construct an agent-facing documentation database containing definitions for thousands of variables across hundreds of datasets. This brings together information that may be scattered across difficult-to-search files or written in unfamiliar languages, so that an agent can retrieve the relevant documentation when advising on a research question, writing code or reviewing it.

The same documentation can guide agents in building synthetic datasets with the structure of the confidential data, so that code can be developed and tested locally before it is transferred to the server. Synthetic data also make it possible to test performance changes, such as switching to DuckDB or Polars where the server supports them, before running them on the real data.

Economic theory and quantitative models
For theoretical work, we will draw on recent examples of AI-assisted research to discuss where agents can contribute and how their reasoning can be checked. For quantitative work, we will discuss how to guide agents when implementing models or adapting existing code, and how to assess whether the implementation follows the intended economics. We will develop skills that help agents use established modelling toolkits and their documentation, such as the sequence-space Jacobian toolkit for heterogeneous-agent models. One potential application is to methods taught in other EUI courses this year, such as Deep Learning for Solving Economic Models and Neural Network Methods for Solving and Estimating Nonlinear Economic Models.

strong>Background
Most participants will already have used AI tools, and many will use Codex or Claude Code. The course concentrates on using these tools more effectively for research rather than on introductory use of chat assistants.

Assessment
The assessment format will be discussed before the course begins.
 

ENROL FOR THIS COURSE