Aircraft from many airlines share a crowded sky, and no pilot decides alone where to go. In controlled airspace, an aircraft says who it is and what it intends, and it moves only as each step is cleared, by a tower that is separate from the pilot. This paper explains that structure in plain language and shows how AI that acts, and that acts on other AI, can be governed the same way.
AI systems are moving from answering questions to taking actions, and increasingly to acting on each other’s outputs. When many independent actors share one environment, safety depends less on how careful each one is and more on how they are coordinated. Aviation addressed this with air traffic control: aircraft declare who they are and what they intend, receive permission one step at a time, and are cleared by a tower that is separate from the pilot. This paper explains those ideas in plain language and describes how the Essence® platform applies the same structure to AI: declared intent, authorization at every step, and a governing layer that is not the AI itself. It also sets out where the analogy stops.
Every day, aircraft from many airlines, flown by pilots who have never met, share the same airspace at hundreds of miles an hour. The pilots are trained and careful. The system is built so that safety does not rest on that alone.
Each pilot sees a small part of the picture: the aircraft nearby, the weather ahead, the instruments in front of them. No pilot can see every other aircraft’s plan. If safety depended on every pilot independently making the right call, one misjudgment could put two aircraft in the same place.
Aviation did not reach its current level of safety only by asking pilots to be more careful. The 1956 mid-air collision over the Grand Canyon, which killed everyone aboard both airliners, is widely cited as a turning point that led the United States to build a system of central coordination. The lesson was structural: in a shared space, safety comes from coordination that no single participant controls.
AI is now entering a shared space of its own. Many AI systems and agents work in the same email accounts, calendars, payment systems, and databases, and in each other’s outputs. Each one sees only its slice, and each one acts quickly. The question is no longer only whether each system is capable. It is what coordinates them.
In controlled airspace, an aircraft does not take off and see what happens. It is identified by a registration and a call sign, and for flights in the busiest airspace a flight plan is filed in advance: where it is going, by what route, at what altitude, and when.
Two things make this work. First, identity is declared and can be checked. The tower knows which aircraft it is talking to. Second, intent is declared before movement. The plan gives the system something to compare against. An aircraft that departs from its plan is visible as a departure, and it needs a new clearance. It does not get forgiven afterward.
Compare a typical AI agent today. It is handed a goal in everyday language and access to some tools, and it chooses its own steps. There is no declared plan to check against. What exists is a log, written as the agent goes and read later, if anyone reads it. The log can show that the agent left its lane. It cannot prevent it.
A pilot does not receive one permission that covers the whole trip. Clearances come in stages: to taxi, to cross a runway, to take off, to climb, to move on to the next controller’s area, and to land. Each one is specific, and each can be changed or withdrawn.
An instruction such as “hold short” tells a pilot to stop before a runway. Being cleared to taxi does not include permission to enter the runway. Authority is narrow: it covers what was stated and nothing beyond it.
This is what a check at the front door lacks. A system that verifies identity once at login and then trusts everything after treats the whole trip as one permission. In the sky, that would be like allowing an aircraft cleared to land to cross any runway it chose on the way.
The same idea applies to handoffs. When an aircraft moves from one controller’s area to the next, the first controller hands it off and the second accepts it and issues instructions of its own. Nothing is assumed to carry over. AI systems that pass work to each other need the same discipline: each handoff is a new check, rather than trust inherited from the link before. Without it, a chain of AI systems becomes a pipeline in which nobody is asking whether the final action is allowed.
A pilot proposes and a controller authorizes. Even the most experienced pilot does not clear themselves to take off.
The reason is the separation of roles. The pilot knows the destination and sees what is nearby. The tower sees the whole airspace, and its job is the safety of all the traffic rather than the progress of one flight. Neither role can do the other’s job well.
In AI terms, the model is the pilot. It proposes what to do, and it is often very good at that. Trouble starts when the same system that wants to act also decides whether it is allowed to. An AI agent that holds its own credentials and permissions is a pilot who also runs the tower, and a model that can be persuaded is a poor choice of tower. Paper 82 described the same idea as a customer at a bank counter who fills out a request slip and never touches the vault. The tower is a second picture of the same rule: the one who asks is not the one who decides.
Air traffic communication uses standard phrases and a habit called read-back. When a controller gives a clearance, the pilot repeats it, and the controller listens for a match. If the two do not match, it is corrected before anything moves.
Aviation adopted this because ordinary conversation leaves room for two people to understand the same sentence differently. A shared vocabulary makes it possible to check that what was meant is what was understood.
Most AI today works from ordinary language and has to work out what a request meant. The same sentence can mean different things in different contexts, and the system settles on one reading. That is a guess, and a check cannot be stronger than the understanding beneath it.
Essence® addresses this with Meaning Coordinates: 256 basic units of meaning, such as create, move, measure, authorize, and restrict. Every request resolves to precise coordinates, so the system does not have to guess. That gives the governing layer something like read-back. The declared intent exists in a precise form, and it can be compared exactly against the action that is about to run. If they do not match, the action does not proceed.
Essence® is designed to play the role of the tower for AI systems that act. Each part of the sky’s structure has a counterpart.
The filed plan. Intent is declared in precise form using Meaning Coordinates, so there is a plan that can be checked.
Registration and call sign. SecuriSync™ establishes who is asking. It confirms identity using keys and a set of registered devices rather than a single credential, so compromising one device is not enough to take over an identity. It issues trust certificates that expire and can be revoked, and a revocation reaches everyone who relies on the certificate immediately. Unlike a registration number, the trust behind it can be withdrawn on the spot.
Clearance at every step. Synergy® is the part that decides what is permitted to run. It makes that decision before each action executes, applies fixed rules, and leaves a record that can be audited. AI can still suggest what to do. Synergy® decides whether it happens.
Rules attached to what is flown over. Nebulo® handles rights. Each piece of data carries its own rules about who may see it, change it, or even know it exists, and those rules are evaluated at the moment of access rather than reused from an earlier answer.
What moves through the system. Everything in Essence® is an Aptiv, a unit that carries what it does, why it does it, and the rules that govern it, all together. Think of an aircraft that carries its own plan and registration wherever it goes. StreamWeave® protects each Aptiv with encryption designed for a post-quantum world. It is also an active defense: it combines many encryption methods, and the combination changes each time an Aptiv is touched, so an attacker who breaks one exchange cannot reuse what was learned on the next. This part has no direct counterpart in the sky.
No analogy is exact. Four differences matter.
Speed and scale. Human controllers can handle only so many aircraft at once. AI systems act far faster and in far greater numbers, so the checking has to be automated. That makes the rules the important part. They must be written, reviewed, and owned by people, and Essence® is designed so that humans remain the authors of what machines are permitted to do.
Who can override the plan. In aviation, a pilot can depart from a clearance, and an aircraft can be hijacked, after which it goes wherever the person in control decides. An Aptiv does not work that way. It executes only what is authorized, so taking hold of one does not enlarge what it is permitted to do. Unlike code-driven or AI-driven approaches, Aptivs are also not controllable by code injection or by persuasion. A pilot can be talked into a deviation, and an AI model can be talked into one too. An Aptiv cannot, because what it may do is set by authorization and not by what it is told. Overriding the plan is not a bypass either. An attempt to override is itself a request, and it is subject to contextual governance: the rules take the circumstances into account, so an override goes ahead only where the rules allow it under those circumstances, and the decision is recorded.
Layers, not a single control. Aviation does not rely on the tower alone. Aircraft carry collision-avoidance equipment as a last resort, and radar and recordings support investigation afterward. Determination does not replace detection. As in Paper 82, the problem is a system that has only the second.
Coverage. A tower governs the aircraft in its airspace. Likewise, the protection described here covers what runs through the governed layer. Systems that are not connected to it are as exposed as they were before, and an AI that holds its own access to them sits outside the design.
MindAptiv is deploying this architecture. The governance claim in this paper rests on how the system is built, which the technical papers describe in detail.
AI that acts is entering a crowded sky. When someone offers you an AI system, ask who is in the tower.
Read the Series → Request Access