Catalog & Cocktails: The Honest, No-BS Data Podcast TAKEAWAY - The way we build agents today is dumb with Vaibhav Gupta
10 snips
Feb 12, 2026 A critique of how agentic AI is built today and why many implementations are fragile. Discussion of LLMs' probabilistic outputs creating tight feedback loops that amplify errors. Comparisons to gaming design for guardrails and feel. Proposal for new abstractions and typed patterns to give AI systems clearer structure and semantics.
AI Snips
Chapters
Transcript
Episode notes
Agents Are Inherently Probabilistic
- AI agents are fundamentally probabilistic and that changes how we should program them.
- Chaining probabilistic outputs compounds uncertainty and makes current agent code fragile and messy.
Learn From Gaming: Guardrails And Feel
- Gaming shows how to handle probabilistic systems via guardrails and feel-based design.
- We need abstractions that capture these benefits similar to TypeScript or async patterns in programming languages.
Define Types Before Using LLMs
- Define structure and types first, then treat LLMs as blob operators that map inputs to typed outputs.
- Connect LLM outputs back to your language's type system to reduce brittleness and improve safety.