AI in practice
The blueprint and the hand
The rule could not tell a label from a subject, because at the level the rule was working, a label and a subject are the same shape.
Agentic work splits into two jobs that look like one job. An agent that is excellent at the first and blind in the second is still worth having, as long as nobody confuses which is which.
What an agent is actually for
The useful picture is not a machine that does the work instead of a person. It is a machine that produces the blueprint, at a level of detail and a speed no person matches, and a person who executes against it and carries the result.
That sounds like a demotion of the human until you look at what each side is good at.
An agent can hold the whole shape of a system in front of it at once. It can enumerate every case, write the plan, and then carry it out across a hundred files without getting bored on file sixty. Boredom is a real failure mode and software does not have it.
What it cannot do is notice that one of those hundred files meant something different from the other ninety nine.
A small failure, last week
Here is one, exactly as it happened, because the useful examples are never the dramatic ones.
A long document had notes scattered through it. Each note was a block that began with a short bold label, then the content. The instruction was ordinary: take the notes out, keep the content, turn each block back into normal text.
The agent wrote a rule. Remove the bold phrase at the start of each block, keep the rest. The rule was correct. It was applied to every block in order, in a fraction of a second, consistently.
And in one of those blocks, the bold phrase at the start was not a label. It was the name of the thing the paragraph was about. The rule removed it, and a paragraph that had explained what a particular substance does now explained what nothing in particular does.
Nothing failed. No error, no exception, no test went red. The document was shorter and still read as English.
The rule could not tell a label from a subject, because at the level the rule was working, a label and a subject are the same shape.
Who found it
A person reading it. Not a checker, not a second pass, not a validation step. Somebody who knew what the paragraph was supposed to say, reading it as a reader, noticing that a sentence had lost its subject.
That took seconds. Finding it from the other direction, by searching the change afterwards, took longer, and only happened because the person said so first.
And there is a detail worth keeping. It was not the only thing the rule had removed. Once somebody knew what kind of mistake to look for, three more were found immediately. The person spotted the symptom. The machine was then genuinely better at finding the rest of the pattern, which is the division working as it should.
So where does the line actually fall
Not between easy work and hard work. The agent does plenty of hard work.
It falls between form and meaning.
A tool that enforces form at the moment work becomes public is worth building and it catches a great deal. It would not have caught this. There is nothing invalid about a sentence missing its subject.
- <b>Form is the agent's.</b> Is this valid, does it parse, is it consistent, does every case follow the rule, does it build. These are checkable without knowing what anything means, and a machine will do them better and more patiently than a person ever will.
- <b>Meaning is the person's.</b> Is this still saying the thing it was supposed to say. Nothing in the shape of the text answers that, which is why no amount of additional automated checking finds it.
What this means for a company introducing agents
The question is not how much to automate. It is where the reading happens, and whether anybody is still doing it.
An organisation that skips the third step will not see failures. It will see output, on time, that parses, builds and ships. The errors it produces are not the kind that announce themselves.
- <b>Let the agent plan and execute at scale.</b> That is what it is for, and holding it back to the pace of human review wastes the thing you are paying for.
- <b>Put a machine gate at the last door</b>, where a mistake becomes public, to catch everything that is checkable by form.
- <b>Then have somebody read the output as what it is,</b> not as a diff. A document read as a document. A page looked at as a page. This is the part that gets dropped first and it is the part that cannot be replaced.
Agentic engineering is not a machine replacing a person. It is a machine that produces the blueprint and does the labour, and a person who still has to read the result and understand what it means, because meaning is the one property that cannot be checked by rule. Decide now who that reader is. If the answer is nobody, that is also a decision.
MK!
AIPathTech Editorial
There is no tracking on this site. No analytics, no cookies, nothing recording where you went or how long you stayed.
But I would like to know if something here was worth your time. If it was, press the heart. It tells me nothing about you. It is a number going up, and it is what tells me to keep writing.
If you want more of this, I am on X. @Queendemetriak