ServiceNow Podcasts The way we build agents today is dumb with Vaibhav Gupta
Feb 12, 2026
Vaibhav Gupta, co-creator of BAML and founder at Boundary ML focused on safe, typed agentic AI tooling. He argues most agent code is fragile and unfamiliar with probabilistic failure. He traces how abstractions evolved in web dev and gaming. He introduces BAML, demoing schema-driven extraction, prompt preview, and static checks to make LLM systems more reliable.
AI Snips
Chapters
Transcript
Episode notes
Agent Code Is Inherently Probabilistic
- Agentic code often fails because it treats probabilistic model outputs like deterministic functions.
- Vaibhav Gupta says developers must rethink application logic to handle inherent model unreliability.
Abstraction Wins—After Many Bad Iterations
- History shows abstraction wins but only after many failed frameworks and new primitives appear.
- Vaibhav compares LLM tooling evolution to jQuery→React→Tailwind as the path to better developer ergonomics.
Game Dev Solves Probabilistic UX Tradeoffs
- Game development solved many probabilistic UX problems by favoring player experience over absolute truth.
- Vaibhav uses latency compensation examples to show choices about which user view to favor.