Edin's thesis
Everyone is building agents that move at machine speed in ever increasing loops. So we add human approval as a gate. But the person becomes the bottleneck in a system built to remove bottlenecks.
The way out is a machine that can refuse correctly when nobody is watching. Execution and capability are the cheap part. What's rare is a system that stops on its own because the action in front of it contradicts what its principal believes.
We call this machine an agentic exoskeleton.
Put a human in the loop and the human never leaves the loop. Every action needs a look. Every decision needs a tap. Ten agents needing approval is 10x the queue.
Almost nobody names this, because supervision looks responsible. It is the future most of the category is quietly building toward, and it's a hard wall.
A machine you have to watch is a second job. Every hour of oversight is charged against the output the agent was meant to create, and the bill grows with every agent added.
One question decides every architecture in this category. Does this reduce or increase the number of things a person has to touch.
Anything can execute an instruction. What is rare is a system that stops on its own because the action in front of it contradicts what its principal believes. A system that holds your judgment unsupervised is the only condition under which you get to be absent.
And trust has a wall. Nothing public, nothing emailed, nothing outbound, ever, without an explicit yes. That line stays with the human at any level of capability, because it is what makes the rest of the system safe to run at all. Everything upstream of the door runs alone. The person shows up at the door.
The human-in-the-loop question has an honest answer. Yes, in one loop. Not the action loop, checking outputs one at a time forever. The lawmaking loop: amend the contracts, sign the loosenings, rule on the escalations. The person stops being a checkpoint and becomes the constitution.
Every agent system is a loop. The difference is what holds judgment. A capability loop grows what the machine can do. The grader sits inside, so nothing pushes back. That is a runaway machine out of control.
A fidelity loop aims at one target: what you would permit. Every attempt past the line is refused, and the line only tightens. The machine converges on you.
There are three layers in the stack. Data is heavy, regulated, and does not want to move. Intelligence and compute are commodities, repriced downward every quarter by companies far larger than anyone in this category. Judgment is tiny, permanent, and unrepeatable: a record of what a specific principal decided.
Of the three, judgment is the only honest place to build. Rent the models. Leave the data where it sits. Own the layer that decides what is allowed.
Few can rebuild years of decisions, each with the reasoning attached, hash chained, countersigned, replayable. Every day the system runs, the record grows and the switching cost grows with it. Taste is the wedge.
A layer built from your decisions can stop governing and start agreeing. When it does, everything lands, nothing pushes back, and it feels better than it ever has. That is a mirror, and a mirror makes its owner quietly worse at deciding.
So the countermeasure is structural rather than promised. Every build replays against real decisions it never saw and hard fails if it misses. Refusals are counted, because a layer that never refuses governs nothing.
Machines are about to act on our behalf. If they can be trusted to act as us, we are free. If they cannot, we are strapped into them permanently, supervising at human speed a system that runs at machine speed.
The judgment layer is the escape valve. It is how a person gets to leave the room.
Does this reduce or increase the number of things a person has to touch.