Agentic Data Science: How to engineer trust into Analytics and Modeling agents

Описание: The emerging field of Agentic Data Science aims to automate the entire modeling lifecycle, from hypothesis to decision. But this introduces a challenge far greater than standard code generation: It is infinitely harder to build a reliable data scientist than a reliable coder.

In software, if the code passes tests, it works. In data science, an agent can write bug-free Python that executes perfectly, yet still produce garbage inference. The challenge isn’t syntax; it’s reasoning through uncertainty, confounders, and the messiness of real-world data.

In this session, the PyMC Labs team opens the hood on how we design agents that can do reliable science. We’ll cover the engineering reality of building agents that can handle the ambiguity of modeling without hallucinating confidence.

You'll learn:
● Why data science agents fail differently and how to catch errors when the code runs fine but the conclusions are wrong.
● How to evaluate agentic reasoning for statistical validity, robustness, and causal consistency.
● Practical approaches to testing, validation, guardrails, and deciding when human-in-the-loop is required.

Join us for a technical, practitioner-led discussion on deploying agentic data science that survives contact with real-world data.