Two Months of Vibe Research with Antigravity

A research agent that works in my folders

Since late November 2025, I have used Antigravity intensively for data analysis, rotating among several Google accounts to avoid quota limits. Unlike an AI isolated in a cloud chat window, it gave me my computer folders as a workspace and functioned as a research partner. The impact was enormous—an eye-opening experience that went beyond the first time I encountered ChatGPT.

The basic workflow of vibe analysis

My core workflow has been as follows. First, I establish the state of the data: what data exist, in what formats, in what volume, and with what degree of completeness. I have the agent prepare a data inventory.

Second, I explore hypotheses. I may already have a hypothesis, but an agent makes it almost free to add a process for exploring candidate hypotheses. This can strengthen the main research hypothesis or reveal genuinely worthwhile ancillary analyses.

Third, I repeat analysis, review, and documentation step by step. I break work into small units—currently distinguishing between studies and steps—conduct analyses divergently, check their validity using a group of chatbots, and then converge on an interim report with tables, figures, and logs.

Divergence without convergence does not create organized context for the agent to use. That makes large-scale research difficult.

The missing ingredient: repetition

I worked on 12 projects, but their topics and methodologies were all different, so there was little experience that could be reused. In an environment with more repetition, I think it would be necessary to systematize repeated work, although I have not yet developed an intuitive understanding of how to do that.

When there were at least similar stages, I copied a .py script from another research folder into the current one and asked the agent, “Can we proceed with this step using this script as a reference?” This made the work much easier. To implement the kind of improvement that comes from accumulated experience, one would need to create skills, but I have not done that yet.

Even with the same agent, performance differences will be enormous depending on how well a person builds rules, workflows, and skills suited to their own work.

The agent codes well, but it has recurring failure patterns

There was no case in which it completely failed to write the code I requested—although the analysis code itself was not especially difficult. Still, several consistent failure patterns stood out.

• Field-name confusion: Problems arose when multiple files used different names, or when names were non-typical. This made me feel that fields should be named well from the beginning.
• Using a function from another script without defining it in the current script: This is a typical context error and is easy to fix.
• Errors from outdated library syntax: Unlike a chatbot, an agent can run code, inspect errors, and debug them, so this was not a major issue by itself. However, it sometimes applied an erroneous method to the entire dataset and produced many null values. Before large-scale processing, it is necessary to validate the procedure on the smallest possible unit.
• Proceeding without checking implicit assumptions: These include units, data structure, and standardization methods. They need to be clarified during data review.

Vibe debugging without reading code directly

In agent-based work, I almost never type code myself, and I read it far less often than before. When an analysis result differs from expectations, or when something feels off in the agent’s reasoning trace, I identify the concern by saying something like, “Check for this problem,” and it usually resolves the issue. Only rarely does vibe debugging fail; in those cases, I have to inspect the code directly and locate the problem.

I worry that junior researchers may be unable to assign work with specific expectations, or may not yet have a sufficiently developed sense that something feels off. They may see a result that looks reasonably acceptable, fail to detect an error, and move to the next step. A new type of risk is that methodological errors that do not appear as syntax errors can quietly pass through.

The human role is to restrain

The agent seems to have a powerful instinct to move forward, and the human researcher’s main role feels like restraining it. It is like the accelerator is pressed by default: I have to keep watching the dashboard and continually apply the brakes.

• “The data transformation is 99% complete and acceptable, so let’s move on.” What?
• “The validation score is 0.99, so it is excellent—let’s move on.” Is that really normal, in common-sense terms?
• “We found a good result while exploring different options, so let’s move on.” Hey, I said EXHAUSTIVE!

A practical review loop

One useful approach was to have the agent produce an .md report for the result of each step, give that report to a group of chatbots—ChatGPT, Claude, and Gemini—and ask them to “audit for scientific rigor.” I then reviewed their feedback and gave it back to the Antigravity agent to close gaps. Repeating this process helped verify the completeness of each stage before moving forward.

Moments that felt like magic

There is an expression, “Feel the AGI.” Two kinds of work genuinely felt magical to me: information integration and contextual understanding, and API-related work.

• Integrating files that were split across several sources, with a separate coding manual.
• Organizing the Excel files, related documents, IRB plan, discussion notes, and poster presentation materials from a project that had been discontinued long ago.
• Starting a study from a single Excel file and seeing the agent infer the full research context from file names, field names, and data values.
• Providing documentation for the Health Insurance Review and Assessment Service API and having the agent download data. It spent hours finding bugs and retrying, overcame rate limiting, and downloaded all the data completely.
• When I wondered about OpenAI TTS voice settings, it used the API to automatically generate sample voices one by one.

What agents do well and poorly

• They do well: integrating widely dispersed data, including not only tabular data but fragmented documents; constructing a coherent story from incomplete data using assumptions; repeatedly trying and moving forward without fatigue when given a goal; and generating various ideas for what to try next based on work completed so far.
• They do poorly: rigorously verifying things one by one, including prerequisites, the completeness of data cleaning, and scientific validity at each stage; suppressing their forward instinct and exhaustively exploring all possible options before moving on; and sensing that something is off, then being willing to trace back several steps—potentially invalidating previous work—to find the problem.

Capabilities needed to collaborate with an agent

• Break work into units small enough for the agent to proceed without difficulty and for the human to review without difficulty; assign the work, review it, and complete that stage.
• Find scientifically insufficient elements—premises, definitions, biases, and omissions—and thoroughly close the gaps.
• Guide the work toward results that fit scientific logic, rather than merely results that look acceptable.
• When something feels off, decide to return several steps, locate the problem, and prevent errors from accumulating further.
• Organize experience into sufficient, reproducible records—data inventories, audit logs, interim reports, rules, workflows, and skills. This is a form of context management.

The greatest risk of vibe analysis

I think the greatest risk is obtaining conclusions that I cannot review. With the agent’s help, analysis may extend slightly beyond the boundary of my existing knowledge. In that case, I can study sufficiently, expand my own boundary, and bring the result into a range that I can verify.

The problem is when the analysis goes beyond even that range. In such cases, a collaborator should be brought in. As AI becomes widespread, I have observed a weakening of collaboration and discussion. I worry that this may create problems for the long-term development of science. We should not swallow something carelessly without even chewing it.

This is a period for embodied learning

Current agents are, of course, imperfect. Even so, I believe that the attitude of “I will do it later, when they are perfect” risks missing the opportunity and falling behind.

What matters is developing a feel for working in new ways as a researcher in preparation for the future, even while the tools remain imperfect. It also means gradually expanding the range of work one can personally review, retaining the drive to improve, avoiding the offloading of everything, and continuing to build one’s own capabilities.

I have posted a YouTube video in the form of a demonstration workshop for those who are interested. I also use rules and workflow settings in Antigravity for a data-scientist role.