Workshop on Using Antigravity as a Research Agent (January 2026)

English summary

This workshop demonstrates Google Antigravity as an autonomous AI agent based Integrated Development Environment for linking data inspection, statistical analysis, visualization, report generation, and manuscript drafting in scientific research. Its central approach is to use an agent that plans and executes multi step workflows while retaining human review of the plan and execution.

Key points

• Antigravity combines a file explorer, integrated terminal, and script editor with an AI agent that can read directories, create folder structures, write and execute Python scripts, and generate documentation. Planning Mode creates an analysis plan and task list before execution, whereas Fast Mode proceeds without detailed prior planning.
• Global rules and reusable workflows can specify raw data preservation, reproducibility requirements, file naming conventions, and default statistical libraries. Examples include data-check for data validation and exploration, imrad-report for an Introduction, Methods, Results, and Discussion manuscript draft, and plan for structured research planning.
• The user reviews the proposed implementation plan and task checklist, adds constraints such as participant de-identification, timestamp based session duration calculations, and avoidance of unnecessary subgrouping, and then permits the analysis to proceed.

Audience

The material is intended for academic researchers, clinical investigators, medical educators, and data scientists who want to use autonomous AI coding agents for empirical research, exploratory data analysis, medical education studies, and scientific manuscript preparation.

Scope and limitations

The main academic case is a workshop demonstration using dialogue data from 19 medical students interacting with LLM based virtual patients and associated survey data. Demonstrated outputs include correlations between turn count and performance, correlations between interview duration and performance, and pre post survey comparisons, but these materials do not establish independent reproducibility or causal relationships. The automated literature workflow may provide a broad overview but does not replace an exhaustive systematic bibliographic search, and agents may generate plausible but inaccurate statistical assumptions, data linkages, or methods. Researchers therefore need to verify package versions, test conditions, handling of missing and zero differences, and significance thresholds. Reported operational limitations include delays during inline Python execution on Windows, inability to directly modify executed notebook cells, and reduced peer dialogue when work is conducted by a single user with AI.

Research context

The workshop is thematically connected to AI supported data science in medical research, reproducible analysis pipelines, and human reviewed use of autonomous coding agents.