Most AI governance frameworks assume a closed-world model: a bounded, controllable agent operating in a predictable environment where behavioral rules can be written and enforced. This assumption is no longer viable. The moment an agent interacts with third-party tools, external APIs, other agents, or multi-step workflows, the environment becomes open, the pathways multiply, and behavioral enforcement collapses. This paper argues that the governance response to open agent ecosystems must shift from enforcement to attestation — from attempts to control what agents do to the production of independently verifiable evidence of what they did, when they did it, who authorized it, and what the exposure window was. Behavioral control is not wrong; it is insufficient. Evidence scales. Enforcement does not. Governance that cannot produce verifiable evidence is not governance at all.
Narnaiezzsshaa Truong (Thu,) studied this question.