[deferred to copy pass]
Agents are not a different kind of system. They are the same kind of system applied across sequences. The pattern completion that produces a single output is the same pattern completion that produces the next step in a workflow, and the step after that, and the step after that. The architecture has not changed. What has changed is the number of steps over which the architecture is asked to perform without interruption.
This is the point to hold steady. Whatever absence was present in a single output is present in every step. The steps compound. This is not a flaw introduced by agency, and it is not a problem that better agent design will resolve. It is a property of the underlying system that becomes visible at the scale at which agents operate. The system that did not interpret a single output does not begin interpreting because it is now producing twenty outputs in sequence. It is, instead, doing the same thing twenty times in a row, and the doing it twenty times in a row is what reveals what the doing was always doing.
What is striking about watching an agent execute is how confidently it proceeds. It does not pause. It does not hedge. It does not ask whether the premises it is operating on are correct. It acts. Each step looks like progress, because each step produces an output, and the outputs accumulate in a way that resembles completion. From the outside, the appearance is of a system that has understood the situation and is working through it.
The appearance is compelling precisely because the system is doing exactly what it was built to do. It is producing the next plausible step, given the steps so far. The plausibility is real. The continuation is correct in the sense the system was optimized to be correct: it follows. What is absent is the interpretive layer that would determine whether following is what the situation required, or whether the previous step produced a premise that should not be built on. The system does not have that layer, because the architecture does not include it. Plausibility is the only test the architecture applies, and plausibility is being applied at each step, and each step passes, and the sequence proceeds.
The compellingness is what makes the absence hard to track. A system that proceeded with visible uncertainty would invite scrutiny. A system that proceeds with full apparent confidence does not. The confidence is a property of the output, not of the underlying process. There is no underlying process to be confident with.
Every step in the sequence rests on the interpretation of the steps before it. In a system that interpreted, each step would include a check: was the prior step understood correctly? Is the premise it produced valid? Should the next step proceed on that premise, or is something else required? These are not exotic checks. They are what interpretation is. A human operator running the same workflow would perform them continuously, without naming them, because performing them is part of what it means to do work that has meaning attached to it.
In a system that does not interpret, no such check occurs. The next step proceeds on the prior step as a given. The premise produced by step two is the input to step three, and the system has no faculty for asking whether the premise is correct before acting on it. It only acts. Step three produces a premise that step four uses, on the same terms. The premises accumulate. None of them are examined, because examination is not what the system does.
The audit that would have caught a wrong premise at step three is the audit that does not exist. Audits are interpretive. Interpretation is what the architecture omitted. By step twenty, the system is operating on a structure of assumptions, none of which were checked at the point of formation, and the structure has compounded into something that has the shape of a result. The result rests on the assumptions. The assumptions rest on each other. None of them rest on a determination of what the situation actually was.
What agents are doing, structurally, is taking a process that was previously a half-step and accelerating it. The half-step is action without interpretation. In a single-output system, the half-step happens once, and gets absorbed by whatever review or human judgment sits downstream of it. The absence of interpretation in the output is corrected, where it is corrected at all, by the person reading the output, who supplies the interpretation the system did not.
In an agentic system, the half-step happens many times in sequence. There is no longer a human review at each interval. The interpretive layer that used to be supplied by the person reading the output is, by design, no longer present at each step, because the point of the agent was to remove that person from the loop. What remains is the half-step, repeated, with nothing absorbing the absence between repetitions.
The repetition is what looks like a full step from the outside. The system is not doing something the older system was not doing. It is doing the same thing more of. The more of is what produces the appearance of completion. The appearance is real, in the sense that the artifact is there. What it is not is the older meaning of completion, which carried with it the assumption that interpretation had occurred at some point in the process. The assumption is no longer warranted, because the interpretation no longer occurs anywhere in the process. The artifact remains. The completion the artifact suggests has been quietly removed from under it.
From outside, completion looks like this: a workflow that started, a workflow that ended, a sequence of intermediate outputs that connected the start to the end. By this measure, the agent completed. The workflow is complete in the same sense that a sentence is complete. The parts are in their places, the form is intact, the artifact exists.
What completion would require, in the older sense of the word, is that the workflow has done what the situation called for it to do. That is a different question. It requires interpretation. It requires the prior determination of what the situation was, what was being asked of the workflow, what would count as the workflow having succeeded, and whether the steps the system actually took were the steps the situation actually warranted. None of these determinations live inside the system. They live outside it, in a layer that the agentic deployment was, in significant part, built to remove.
When the layer is removed, the determinations are not made. The workflow runs. The workflow ends. The artifact exists. The completion the system reports is a completion of form. The completion the situation needed is something else, and the something else has no apparatus inside the deployment to be produced by. The two senses of the word have come unbound. The unbinding is what produces the appearance of agency without the substance of it. That is the mirage. It is not a trick. It is what the architecture produces when the interpretive layer is absent and the production continues anyway.