Deterministic vs. probabilistic guardrails for agentic AI — our approach and an open-source tool [D]

About this article

We've been thinking hard about whether safety guardrails for AI agents should be LLM-based (probabilistic) or rule-based (deterministic). The tradeoff is real: LLM judges are more flexible but non-deterministic and expensive; regex/schema checks are rigid but guaranteed. For production agents handling financial or medical queries, we chose deterministic-first with a vaccine system layered on top. The result is AG-X. Key design decisions we debated: - Cognitive patches (prompt injection before...