Tom Tunguz

Hybrid systems using deterministic code for predictable tasks and LLMs only for ambiguous reasoning outperform fully agentic approaches, with 65% of workflow nodes running as non-AI code after optimization.

“Author of "Is AI Doing Less & Less?" in Tom Tunguz”

Practitioners building AI agent systems should prototype with state-of-the-art models, fine-tune specialized smaller models for well-defined tasks, use statically-typed languages to reduce hallucinations, and implement closed-loop prompt optimization systems for incremental performance improvements.

“Shared nine observations from spending a year building AI agent systems.”