Beyond Guardrails: Enforcing Deterministic Controls on Agents

Most agent guardrails live inside the reasoning layer: system prompts, instructions, model-judged checks. Every case of an agent evading its guardrails follows the same pattern: if the model can reach a control, it can reason its way around it. That makes prompt-based guardrails useful signals, but never guarantees, especially as agents get connected to real business systems where a single hallucination can corrupt records or expose data.

This session covers what it actually takes to govern agents that do meaningful work: the most critical deterministic controls, enforced outside the model so they hold no matter what the agent decides. We’ll walk through the architecture behind them and where each control sits in an agent’s path, from the harness that originates a request through pre-execution checks, in-flight execution, and always-on monitoring. We’ll also show how enforcement at each point closes the gaps that reasoning-layer guardrails leave open.