Vibe Research Using Agents (2026.2)

English summary

This presentation explains how passive large language model chatbots can be extended into autonomous AI agents that plan, use tools, execute tasks, receive feedback, and iterate, and how such agents may be applied to scientific research. Drawing on the presenter's two-day evaluation of 12 projects using Google Antigravity, it describes practical uses, failure modes, and the verification competencies required of researchers.

Key points

• Traditional chatbots mainly generate one-off responses for closed-world tasks and do not directly access external environments, file systems, or persistent context. AI agents are configured to plan and perform long-term, multistep tasks in open environments, execute code, use APIs and file systems, and produce verifiable deliverables.
• The ReAct structure is described as a repeated sequence of thought, tool action, observation, and subsequent thought. A three-layer agent architecture is presented as foundational agentic reasoning, self-evolving reasoning using feedback and memory, and collective reasoning among multiple specialized sub-agents.
• Vibe Research is presented as an extension of Vibe Coding in which researchers use natural-language instructions for data analysis and research tasks. Its central workflow includes creating a data inventory and checking integrity, exploring candidate hypotheses, and iterating through analysis, review, and interim reporting. Multiple analytical paths are explored, critiques from several language models are considered, and an interim report is used to consolidate findings and reset the working context.

Audience

The intended audience includes academic researchers, biomedical and clinical researchers, pharmacologists, computational biologists, and graduate students seeking to understand AI-assisted research workflows.

Scope and limitations

The evidence consists primarily of the presentation and slide content from February 24, 2026, together with the presenter's early observations from evaluating 12 projects over two days with Google Antigravity. The reported capabilities and failure modes therefore reflect limited early experience and emerging frameworks, not a general performance evaluation or a controlled comparative study. Agents may proceed without exhaustive validation, fail to correct early methodological errors, confuse variable names, produce scope errors, use deprecated syntax, or act on unverified assumptions. They may also generate plausible conclusions that are difficult for a researcher to verify critically.

Research context

The content is thematically connected to medical AI and data science through its focus on designing workflows for autonomous agents in research data management, iterative analysis, and documentation, while preserving human verification. Its emphasis on data integrity checks, context management, methodological review, and researcher training is particularly relevant to this area.