AI Agents and Prompt Engineering in Econometric Coding

We study how large language models write econometric code. On 40 tasks in Stata, R, and Python, we vary the prompt (with or without an example) and the degree of agency (from a chatbot that only writes the script to an agent that also runs and revises it). For Claude Sonnet 4.6 and GPT-5.6 Luna, the agent raises task success from 79 to 97 percent. Agent runs cost 3.1 to 5.0 times as much as chatbot runs. The example and the software matter more under the chatbot than under the agent, which removes most failures but only some wrong estimates