AIエージェントと組んだら、データサイエンスプロジェクトはどう変わる？実験してみた（後半戦）The second half of a hands-on experiment exploring how AI agents reshape data…

The rise of autonomous AI agents is prompting data scientists to reconsider how routine analytical work gets done, and a recent hands-on experiment published on Qiita examines exactly that shift. Continuing from an earlier installment, the second half of the write-up focuses on practical workflow changes and productivity effects when an AI agent is embedded directly into a data science project, offering a grounded look at where such tools help and where they still fall short.

Unlike a conventional chat assistant that simply answers questions, an AI agent is designed to take a goal, break it into steps, execute actions, observe the results, and iterate. In a data science context this can mean loading a dataset, running exploratory analysis, generating visualizations, testing hypotheses, and drafting model code with limited step-by-step prompting from the analyst. The experiment appears to treat the agent as a collaborator that handles mechanical portions of the pipeline, freeing the human to focus on framing problems and interpreting outcomes.

The reported impact centers on the early and middle stages of a typical project. Tasks such as data cleaning, schema inspection, boilerplate feature engineering, and first-pass charting are often repetitive and time-consuming, and these are the areas where agent collaboration is likely to yield the clearest gains. By generating and running code in a loop, the agent can compress work that might otherwise take hours into a shorter cycle, while surfacing anomalies or distribution issues that a human might overlook on a first pass. This mirrors the broader industry pattern in which AI assistance tends to accelerate well-defined, structured subtasks more reliably than open-ended reasoning.

At the same time, the experiment underscores that oversight remains essential. Agents can produce plausible-looking analysis that is subtly wrong, misinterpret ambiguous column names, or make unstated assumptions about data types and missing values. Statistical judgment, domain knowledge, and validation of results still rest with the human practitioner. The productivity benefit therefore appears to depend heavily on how clearly the analyst specifies the goal and how carefully they review each intermediate step, rather than on handing over the project wholesale.

This work fits into a fast-moving ecosystem of tooling. GitHub Copilot, the category under which this piece is filed, has expanded from inline code completion toward more agentic behavior that can plan and edit across files. Comparable capabilities appear in tools such as Cursor, Claude Code, and OpenAI's coding-oriented offerings, while frameworks like LangChain, LlamaIndex, and AutoGen provide the scaffolding for building agents that call functions, query databases, and chain reasoning steps. For data science specifically, the notebook environment remains a natural home, and integrations that let agents operate inside Jupyter or similar interfaces reduce the friction of moving between human and machine work.

Several prerequisite concepts help explain why results vary. Agents rely on tool use, meaning the model is given access to defined functions or code execution rather than only producing text. They also depend on context management, since large datasets cannot be fed wholesale into a model with a limited context window, so effective setups sample data, pass summaries, or let the agent write code that processes full data outside the model. Retrieval-augmented approaches and structured prompting further shape how reliably an agent grounds its actions in the actual dataset rather than in generic assumptions.

The experiment's value lies less in any single benchmark and more in its qualitative account of a real workflow. It suggests that agent collaboration changes the rhythm of analysis, shifting the analyst toward supervision, specification, and review, and away from typing routine code. Whether this translates into durable productivity gains is likely to depend on the complexity of the task, the quality of the data, and the discipline of the review process.

For readers weighing similar adoption, the takeaways are pragmatic. Start with bounded, verifiable tasks, keep a human in the loop for statistical and business decisions, and treat generated code as a draft to be checked rather than a finished result. As agent tooling continues to mature, these hands-on reports offer a useful, tempered counterweight to broader claims, showing both the concrete time savings and the persistent need for expert judgment in data science work.