By Jon M. Watson, Execution Boundary Governance Architect & AI Strategy Leader
The shift toward action architecture is real progress. Architects are rightly arguing that governance must move from policy documents and review boards into the actual structure of systems — where decisions become consequences. This is necessary. It is not yet sufficient.
The missing variable is the hard physics constraint that most current proposals have yet to confront. At machine speed, an agent can execute a decision and produce irreversible state change in roughly ten microseconds. Any governance mechanism that still requires interpretation, human judgment, policy evaluation, or review processes operating on human timescales cannot intervene before that state change has already occurred. By the time those mechanisms engage, thousands of additional decisions have already executed. This is not a tooling gap or a maturity gap. It is a structural limit on what can bind at the point where consequence forms.
Faced with this speed, the dominant instinct remains additive. Organizations respond to agentic risk by layering in more monitoring, additional AI oversight over AI agents, and more complex governance stacks. This is the default response inherited from cloud-era governance — systems and processes designed for human-speed transaction times and interpretive decision points. It is understandable. It is also structurally mismatched to the constraint.
Many current proposals for AI governance compound the problem by leaning on LLM-based policy evaluation. These approaches are inherently slow (subject to inference and token generation latency), token-heavy, and probabilistic. They introduce interpretation and non-determinism at the exact moment when governance must be fast, binary, and reliable. Larger and more sophisticated systems built on these foundations still cannot close the latency gap or remove the interpretation problem at the point where state change becomes irreversible. The solution to a governance problem created by machine speed is not more layers and more intelligence of the same kind. It is the opposite: the smallest possible structure that can operate inside the physics window without requiring interpretation or probabilistic evaluation.
What the physics requires is a mechanism that is minimal enough to sit in the hot path, immutable so it cannot be altered by the running system, binary so there is no interpretive layer at runtime, and positioned exactly at the commit boundary — the last decision before state becomes irreversible. This is the minimum structure capable of making governance bind before consequence occurs. The minimal kernel is not offered as a complete governance system. It is the necessary nucleus: the non-AI anchor around which higher-level architecture, policy, and oversight can actually function at machine speed.
Human oversight does not disappear under this approach. It moves upstream. Humans retain responsibility for deriving the conditions that should govern machine action, designing the boundaries and escalation paths, and handling genuine exceptions. What must be accepted is that humans cannot remain in the runtime loop at the commit boundary itself. Latency makes direct human judgment or review at that point impossible. Any attempt to keep interpretive control inside that window creates a structural gap between intention and consequence. The kernel is what allows humans to exercise real authority without maintaining the illusion that they can operate at the speed where state now changes.
The more persistent barrier is not technical. It is that many of the humans responsible for governance — architects, risk leaders, regulators, and executives — are not yet willing to relinquish direct control at the exact point where consequence forms. As long as that reluctance remains, even the strongest versions of action architecture will continue to operate one layer removed from where decisions become irreversible. The minimal kernel at the commit boundary is one concrete expression of what accepting that constraint actually requires.