Why AI Safety Should Be Structural, Not Just Rule-Based

A developer argues that effective AI agent safety relies on structural constraints rather than solely on in-context rules or action-level guardrails. The author illustrates this by running an HTML renderer in a sandboxed environment with strict resource limits, ensuring unsafe behavior is prevented by design rather than by the agent's own judgment. Two self-imposed practices reinforce this approach: using a throwaway browser profile by default instead of a logged-in one, and passing secrets via file paths rather than exposing them in the conversation. The author acknowledges these structural choices come with real inconveniences but notes they have repeatedly overridden the more convenient option. The piece closes by inviting others to share what they have made structurally impossible for their agents, rather than simply instructed them to avoid.