Catalog & Cocktails: The Honest, No-BS Data Podcast The way we build agents today is dumb with Vaibhav Gupta
13 snips
Feb 12, 2026 Vaibhav Gupta, co-creator of BAML and systems engineer who builds tooling and languages for reliable agentic and probabilistic code. He argues most agent code is ugly and fragile. He compares abstraction waves from web dev to LLMs. He explains why abstracting failure matters, demos typed prompts and schema enforcement, and shows how languages and tooling can make probabilistic systems robust.
AI Snips
Chapters
Transcript
Episode notes
Games Favor Perceived Correctness Over Absolute Truth
- Game developers solved many probabilistic UX trade-offs by favoring perceived correctness.
- Developers sometimes choose which client to 'lie' to to preserve user experience.
Treat LM Outputs As Typed Contracts
- Enforce schemas and business rules around LM outputs and treat violations as exceptions.
- Convert or null out invalid values so downstream code depends on clean, typed data.
LLMs Improve But Failures Persist
- Model improvements lower but never eliminate failure rates, so systems must tolerate residual errors.
- Any non-zero error rate requires exception design for critical workflows like finance.