The AI Native Dev - from Copilot today to AI Native Software Development tomorrow How New Libraries Saw a 50% Improvement | Maria Gorinova
5 snips
Dec 9, 2025 Maria Gorinova, a member of the AI engineering team at Tessl and an expert in data science, dives into the challenges of using coding agents. She discusses how structured context can enhance reliability and improve adherence to abstractions by 35-50%. The conversation highlights common failure modes like hallucinations and version confusion. Maria also introduces Tessl's innovative registry tiles, which streamline library interactions and boost agent performance. The insights shed light on the evolving landscape of AI-driven software development.
AI Snips
Chapters
Transcript
Episode notes
Abstractions Must Be Respected
- Software development relies on layered abstractions that agents should use rather than regenerate.
- Forcing agents to rewrite established libraries wastes time and loses optimizations embedded in those libraries.
Treat Agents As Probabilistic Systems
- Everything about AI agents is probabilistic rather than deterministic.
- Use population-level evals to make statistical judgments about agent behavior instead of anecdotes.
Measure Library Usage With Targeted Evals
- Evaluate library usage by generating tasks and explicit criteria tied to real APIs.
- Use an agent-as-judge to score solutions against those criteria to measure abstraction adherence.