What Closes the Loop?

The big question right now seems to be whether we can set the agents free and just keep pressing Continue. Let them build, test and fix themselves until it is done, no one at the keyboard. It is an exciting question. But I think it is the wrong one. The right one is simpler and more uncomfortable. What actually closes the loop?

Same agents, different results

I run agents every day, so I am saying this from the inside. Set a swarm of them loose on a large unfamiliar system and ask them to find problems, and they come back with confidence. Line numbers, priorities, serious wording. Half of it does not hold when you check it. Give the same agents a small bounded change right before a release and they find the bug you missed yourself.

The difference was not the quality of the agents. It was the same agents. The difference was how verifiable the task was. On the small and bounded, they delivered. On the large and unknown, they started guessing. And a guess presented together with a line number suddenly feels like a fact.

Proof, not claims

That is where the loop closes or does not. Not in how smart the model is, but in whether the task could be proven and whether you demanded the proof. An agent that says something is wrong has given you a claim. An agent that shows you the failing test has given you a proof. There is a difference. And the difference is the whole job.

So I do not ask for a judgment. I ask for proof. A test that goes from red to green. A line I can read and understand. Something I can inspect, not something I am meant to trust. Demand proof, not claims, and a confident guess stops looking like an answer.

An AI checking an AI

"But have another agent review the first one." Sure, do that. Several agents checking the same change catch more than one. But be clear about what it is. An AI checking an AI closes the loop syntactically. It checks that the build matches what was defined. It cannot check what you forgot to define, because it shares the same blind spot. Two prediction systems that agree are still just agreeing. They own nothing.

The last gate is a human for one reason. Not because the human is flawless. But because someone has to be able to tell the difference between a proof and a well-worded claim. And someone has to own the outcome when it turns out the proof was not enough.

How I close it

There is no magic to it and it is not against agents. Plan before the agent starts. Decide what done means and write it down. Give it tests. Demand proof, not claims. And let someone other than the one who wrote it review the result. Verification is not what slows the work down. Verification is what lets you dare to trust it.

The question is not whether, but what

The line for how much we can trust the models keeps moving. But the principle does not move. The question is not whether the human should be in the loop. The question is what closes it.

And what closes a loop is not more automation. It is proof. And someone who can tell the difference between proof and a claim.

See also: The Layer Nobody Draws (series 40), The Checkpoint Is the Job (series 11) and You Can't Validate What You Don't Understand (series 6).