Two-month experience with antigravity vibe research (January 2026)

English summary

This video describes a two-month experience, from November 18, 2025 to January 27, 2026, using Google's AI coding and research agent Antigravity for Vibe Research and Vibe Analysis. The speaker reports conducting 12 exploratory projects and submitting four papers during that period, and presents a workflow in which high-level research intentions and intuition are communicated to an agent that generates and runs analysis code.

Key points

• The basic workflow consists of creating a data inventory, exploring hypotheses, and conducting stepwise analysis, review, and convergence. After documenting data structure, missingness, and distributions, the agent proposes hypotheses and ancillary analyses, while projects are divided into studies and smaller steps for exploratory work.
• The researcher exported interim outputs to other large language models, including ChatGPT, Claude, and Gemini, for scientific rigor audits. Consolidated critiques were then returned to Antigravity to address methodological gaps.
• Direct code writing and reading were substantially reduced, with Vibe Debugging centered on natural-language requests for rechecking when anomalies appeared. Direct code inspection was still used when language-based guidance did not resolve an issue.

Audience

The content is intended for academic researchers, data scientists, clinical researchers, and software engineers using LLMs and autonomous coding or research agents in data analysis and scientific workflows.

Scope and limitations

This summary is based only on the supplied video and presentation notes. The reported project count, paper submissions, agent capabilities, and failure modes reflect the speaker's experience and observations rather than independent validation or a controlled comparative evaluation. The video reports recurring problems involving field-name and schema confusion, loss of context between sessions, deprecated syntax and version mismatches, and premature progression without adequate validation. It also emphasizes that subtle methodological errors may propagate through later analyses and that users with limited domain knowledge may be unable to determine whether conclusions have reached a verifiable boundary.

Research context

This topic provides a thematic connection to Sangzin Ahn's portfolio through medical AI and data science, particularly research automation, analytical validation, and collaboration between human researchers and AI agents.