Agent factories will not stay in software

Agent factories will leave the IT department, and the first business functions to run that way will be the ones whose work is already written down. A software team that runs as a factory no longer has one person driving one agent. Work enters a queue, agents carry it through stages, and people stand at a few fixed points where something has to be decided. That shape does not depend on the work being code. It depends on three properties the work has to have first.

This matters to a programme like ours because an agent factory is the smallest honest version of what we study. It is a closed operating loop, bounded to one function, with humans still holding the decisions that carry liability. Most firms will meet the autonomous enterprise this way, one loop at a time, long before anyone talks about a business that runs itself.

How software teams got there

The path is easy to see in hindsight. First a chat window next to the work, with answers pasted in by hand. Then suggestions inside the editor. Then an agent on one engineer’s laptop, steered step by step. Then the factory: a ticket enters a loop, agents triage, describe, change and check it, and a person approves the description, confirms the check and agrees to the release.

The gap between the last two steps is larger than it looks. An agent on a laptop is an extension of one person. A factory turns that private habit into a process the whole team can see, measure and change. The same four steps apply to a finance team that asks a chat for a spreadsheet formula today and could run invoice matching as a loop tomorrow.

The condition is not the model

Software went first for a plain reason. Its work has long met three conditions the rest of a firm often does not.

The input is written down. A ticket, a change description, a version history. An agent has something to read, and so does the next person.

The output can be checked without a meeting. Tests pass or fail. The application starts or it does not. Somebody can say “done” and point at evidence.

Every change leaves a trace. Who changed what, when and why can be reconstructed a year later. That is the property we have argued for in a log nobody can replay is not a ledger, and a factory without it cannot be audited, only trusted.

Now look at a month-end close. Half of it lives in a workbook one person understands. The adjustments are explained in an email thread. Whether the close is right is settled by whoever signs it, from memory. An agent cannot run that process, and neither could a new hire. The first step towards a factory in finance is not buying an agent. It is writing the work down so that input, output and trace exist outside anyone’s head.

Two jobs instead of one

In a working factory, each person has two jobs. The first is the work itself: the campaign, the close, the purchase order. The second is the loop that produces it: correcting the instructions agents follow, choosing which model handles which stage, measuring how much passed through without a correction and what it cost.

On our homepage these are the operate and evolve loops. Operate does the work; evolve changes the configuration that did it. In a factory the evolve loop is not autonomous. A person proposes or approves every change to the instructions, and the change is kept as a version that can be rolled back. The controller still signs the month. Her real work becomes a close that sends her fewer exceptions each month, and the record of why.

Where a factory does harm

Two failure modes already show up in software teams, and both will travel.

The lights-out factory. Some vendors promise loops with no person in them. In software that is unrealistic; in finance and sales it is dangerous. An agent that sends an upsell offer to a customer with an open complaint, or posts a quarter-end correction on its own, has not saved work. It has created work for the legal team. The reason is not technical. As we wrote in the last human in the loop is a signature, somebody has to be answerable for each commitment, and the law puts that somebody among people. The agent analyses, prepares and proposes. A person decides, signs and answers for it.

Work that suddenly becomes public. On a laptop, how someone talks to an agent is their own business. In a factory every instruction and correction is visible, and the cost of each person’s work is easy to compute. If management turns those numbers into a ranking of people, the best ones leave first. Numbers from the factory exist to improve the factory.

How to tell it is working

Vanity metric	Outcome metric
documents, emails and reports the agents produced	how many times a person touched one task before it closed
number of staff with an agent licence	share of tasks that passed through the loop with no correction after approval
tokens consumed per department	cost of one closed task, and its trend month on month
“the agents now handle the close”	exceptions that reached a person, and how many of them were real

Be fair about what a loop can change. A campaign can fail because of price. A close can slip because a supplier never sent an invoice. A factory shortens the road to the truth; it does not change the truth.

A test you can run this week

Pick one task in one function that comes back every week and has a clear start and end: a sales report for the board, a product description for the shop, a bank reconciliation. Write it down as if you were handing it to someone on their first day: what arrives, what has to leave, how to tell it is right, and who approves it. If it does not fit on one page, an agent will not do it. A new colleague would struggle too.

That leaves a choice. You can wait until agent factories mature inside IT and only then look at the rest of the firm. Or you can start writing down the work in finance, purchasing and marketing now, so that the day the tools are ready, the loop has something to read and somebody named to sign.