I've seen a lot of talk about "Agentic Data Science" where multiple AI Agents & Automation work together—one for cleaning, one for EDA, and one for model selection. Is this the future of the field, or does it just create more layers of debugging for the average Data Scientist?
3 answers
Multi-agent systems are fantastic for rapid prototyping, but I still prefer a manual touch for final model fine-tuning.
We’ve been experimenting with this in our R&D lab throughout late 2024. The benefit of using AI Agents & Automation for the initial data pipeline is the sheer speed of iteration. We can now run 50 different feature engineering experiments in the time it used to take to run five. However, the debugging is indeed a nightmare if you don't have good observability. You need to treat the agents like junior developers—review their "code" and their logic at every step. It hasn't replaced our data scientists, but it has definitely changed our role to more of an "Architect" or "Reviewer."
How are you handling the version control for the code that these agents are generating during the EDA phase?
Michael, the agents are programmed to commit every successful iteration to a separate "agent-branch" in Git. This allows us to use standard CI/CD tools to review the code they produce. By treating AI Agents & Automation as part of our Git workflow, we maintain a clear history of how our models evolved and ensure that no "ghost code" makes it into production.
I'm with you, Justin. Agents get us to the 80% mark incredibly fast, but that last 20% of precision still needs human expertise.