Why I build agents

For as long as there has been software, people have had to operate it.

We built better and better tools, and every one of them came with a job attached. Someone has to fill in the form. Someone has to run the report. Someone has to read the dashboard, notice the number that moved, and decide what to do about it. The software holds the data. The person does the work.

Agents are the first software that turns that around. An agent does the work and asks people only for judgment.

That is the whole reason I build them.

Most work is reading and deciding

Look closely at what most people in a company actually do all day. A surprising amount of it is the same loop: read something, compare it with something else, decide what to do, tell someone.

The month-end close is that loop, run by a team, under a deadline. Pull the balances, match them against the statements, chase the differences, explain the ones that remain. The hard part is not the arithmetic. It is the hundreds of small checks that agree, which have to be done before anyone can get to the few that do not.

An inbox is that loop. So is a support queue. So is the list of things everyone agreed to at the end of a meeting, which nobody writes down and nobody follows up on.

None of this is creative work. It is careful work. And it is exactly the kind of work software has always been bad at, because it needs a little judgment at every step.

Reports do not learn

The usual answer has been a report. Put the data on a screen, sort it, colour the bad rows red, and let a person do the rest.

The trouble with a report is that it is the same size every day. It shows you the row you explained yesterday. It shows you the exception that is always there and always fine. It cannot tell the difference between a problem and a pattern you already know about, because it does not know anything. So people learn to skim, and then they learn to skip, and the one row that mattered goes unread.

An agent that learns gets smaller. Every time a person explains something, the agent keeps the answer and stops asking. Every time a person makes a decision, the agent can make the next one like it, or at least put it in front of the right person with the evidence attached. Over time, what reaches a human is only what needs a human.

That is the difference that matters to me. Not that the machine is faster. That the pile shrinks.

What it looks like

Some examples, none of them exotic.

An inbox that triages itself. It reads every message, answers the routine ones, files the noise, and shows you the three that actually need you, with a draft reply already written.

A support queue whose easy half resolves itself. The password resets and the “where is my order” questions get handled. The person on the queue spends their day on the problems that need a person.

The follow-up nobody sends. The meeting ends, and the agent writes down who agreed to what, sends the note, and checks back next week on the items that did not move.

A hospital discharge summary written while the doctor is still in the room. The doctor reads it, corrects one line, and signs it, instead of writing it from memory at the end of a long shift.

In every one of these, the agent does the work. The person decides.

What makes an agent worth trusting

None of this works if people do not trust the output. And trust is not a feeling. It is a set of habits you can build into the system.

It shows its evidence. Every claim points at where it came from: this line, that record, this message. If you want to check, you can, in one step.

It asks instead of guessing. When the data cannot answer a question, a good agent says so and asks. A confident wrong answer does more damage than an honest question.

It remembers. An answer given once should never be asked for again. An agent that keeps asking the same thing is a report with extra steps.

A human signs off on what matters. The agent can prepare the decision, rank it, and explain it. For anything with real consequences, a person still makes the call. That is not a limitation. That is the design.

Get those four right and people stop checking everything, because they have checked enough to know they do not need to.

Why I am doing this

Every wave of machines was supposed to end work. None of them did. They moved it up a layer.

The loom did not end clothing. It moved people from weaving to running mills. The spreadsheet did not end accounting. It moved accountants from adding columns to asking what the columns meant. Each time, the routine part went to the machine and the judgment stayed with the person, and there turned out to be more judgment to go around than anyone expected.

This is the layer moving again. The reading, matching, chasing and summarising that fills most working days is about to be done by software that does the work instead of waiting to be operated. What is left for people is the part that was always the point: deciding.

I want to be one of the people building that layer. Not writing about it, building it, sitting next to the people whose work it changes and making sure it actually helps them. That is the job. I like it a lot.