MedAgentGym

Fits code-writing agents for clinical data analysis and biomedical research support. Borrow the Docker sandbox with request-info, code and debug actions, the 12-source task taxonomy, and the split between exact-match coding tasks and accuracy-scored ML tasks.

What it asks the agent to do

• write and run code to answer a question over EHR tables
• perform a clinical calculation in code
• solve a bioinformatics or biomedical data-science coding task
• train or run a prediction model on clinical data
• debug from execution feedback inside a sandbox

What to borrow from it

sandbox designbaselinetaxonomy

Caveats

Mainly a training environment (ICLR 2026) rather than a fixed benchmark; most preprocessed data is gated behind a PhysioNet data use agreement and an email request, and the MIMIC-III, eICU and EHRSHOT portions are excluded from the public download; no license file in the repo; task count is instances across train and test, not a curated test set.

Details

Links

• Paper
• Github