The Data Exchange with Ben Lorica Coding Agents Meet Data Science
22 snips
Mar 26, 2026 Mikio Braun, Senior Principal Applied Scientist at Zalando who builds AI-powered developer tools, discusses coding agents applied to data science workflows. He covers practical limits like unvetted data and timeouts. They explore team-level effects: faster velocity, testing and review bottlenecks, and how agents change collaboration and skill needs.
AI Snips
Chapters
Transcript
Episode notes
Training Data Limits Make Agents Weak For Data Science
- Coding agents were trained on code repositories, so they lack deep data-science context like domain nuances and evaluation practices.
- Ben Lorica noted GitHub training data contains code but not explanations about data cleanliness or evaluation methodology.
Provide Context Stores And Structured Models First
- Build a data-science context layer and leverage structured-data foundation models before asking agents to iterate.
- Ben recommends solutions like Kumo for structured-data models and a context filesystem (DEX) to capture prior team work.
Agents Turn Individuals Into High Velocity Teams
- When every developer uses an agent, each person effectively becomes a small team and velocity increases dramatically.
- Ben Lorica argues this makes code review and testing the new bottlenecks and higher-order human review skills more valuable.