The firm of one
Four AI agents, one manager and the oldest problem in business
Once agents act rather than answer, a single person is running an organisation, and organisations need management, not better prompts.
Essays · first essay planned for December 2026
When work becomes cheap, accountable judgment becomes the scarce resource: a firm of one, run with four agents for a year, shows what a manager must still decide, approve, pay for and remember — and where that lesson stops at the factory gate.
The series returns to four questions: who may decide what without asking, what the work costs and what the bill hides, how it breaks and who notices, and what the organisation knows when the worker forgets. Each essay opens with a management problem and closes with a decision. None closes with a tool.
It is not a build log, and it does not claim that a private laboratory proves anything about a factory. It ends with where the lesson stops.
The firm of one is a private project, run in private time on private equipment. It touches no employer data.
One email per essay, sent by me. Write a line and I will add you; write another and I will take you off.
Four AI agents, one manager and the oldest problem in business
Once agents act rather than answer, a single person is running an organisation, and organisations need management, not better prompts.
Machine labour is nearly free. The bill has moved elsewhere
The marginal cost of machine work has collapsed, so cost control moves from price per unit to capacity, tiering and attention.
I named my agents. The research says that was a mistake
An agent is equipment with an owner, not a colleague with a role; the framing changes how carefully humans check.
This part of the series will be announced separately.
A year of failures in a company with no employees
Agent systems fail like organisations - at interfaces and through defaults - rather than like software.
Why measurement rots, and what a number needs to stay alive
A metric without an owner and a decision attached decays, however good the cockpit.
Agents remember nothing. The organisation has to
When labour is stateless, institutional memory is the asset, and it has to be built deliberately.
A small business, a weekly report and an approval button
The test of an agent system is a customer who does not care how it works.
Sovereignty is not owning everything. It is being able to leave
Sovereignty is a sourcing question; the measure is switching cost, not ownership.
This part of the series will be announced separately.
Method
Research and early drafts are prepared with the AI agents the series describes. The numbers are checked against logs and invoices and are given with their period and date. Estimates are labelled as estimates. The openings, the conclusions and every opinion are mine.
Incidents are published only after they have been fixed. No essay names a host, a provider or a product, which costs some colour and spares some trouble.
The essays can be adapted for management and trade publications, and each one makes a talk.