Super Data Science: ML & AI Podcast with Jon Krohn 959: Building Agents 101: Design Patterns, Evals and Optimization (with Sinan Ozdemir)
47 snips
Jan 20, 2026 
Sinan Ozdemir, an AI entrepreneur and bestselling author of Building Agentic AI, dives into the nuances of agentic AI versus traditional workflows. He discusses how to evaluate AI models effectively beyond just accuracy and shares insights on the right types of models for specific tasks. Sinan also highlights the importance of context windows in agent systems, the trade-offs between precision and recall, and the implications of hybrid workflows. Expect surprising findings on reasoning mechanics and practical guidance for deploying agentic AI.
AI Snips
Chapters
Books
Transcript
Episode notes
Agents vs Deterministic Workflows
- Agents let an LLM choose tools and the order of actions, unlike deterministic workflows.
- Agency introduces unpredictable branching that requires different design and auditing.
Start By Mapping The Existing Process
- Map the existing human process first to decide if you need an agent or a workflow.
- Count conditionals: many branching decisions point toward agents or hybrids, few point toward workflows.
Rough Model Size Tiers For Tasks
- Parameter count gives a rough capability signal but bigger isn't always better for specific tasks.
- Sinan frames tiers: <10B (small), 10–100B (medium), 100B+ (large) to guide model choice and cost trade-offs.
