2026–present · Research in progress
Health Agent Sandbox
A research project on simulation environments for long-term health-agent decisions.
Research question
How can we evaluate health agents when their decisions affect what happens next, rather than only the answer to a fixed question?
Project scope
Health Agent Sandbox explores how to evaluate an agent when its decisions change a patient’s future course. The project is currently at the research design stage, with a specification for reproducible trajectories, alternative decision paths, and outcome-based evaluation.
My work
I am working on the research design and evaluation specification, with a focus on long-term decisions, reproducible patient simulation, and comparisons between alternative decision paths.
Relation to HAEnv
HAEnv provides an existing pipeline for generating and evaluating synthetic patient scenarios. Health Agent Sandbox is a separate project investigating an environment where an agent’s actions change the subsequent patient trajectory.
Current status
Research design in progress. Implementation and validation are planned.