A thesis on companies
composed of agents.
FTE, n. One person’s full working week.
Here, none is spent on work a machine could do.
Who works here?
Every era redraws the company
around its newest tool.
This one is drawn around agents.
An interactive sculpture by zer0fte. Every thread is an agent. The blue one is a person.
The line closes into a loop that learns.
A thesis
by zer0fte
First, we hired
a person for every task.
Then, we bought
software for every person.
Now, we compose
agents for every step.
Eight chapters on how a company gets made.
Start with the question. Follow the work.
How many people does a company need? Zero?
Scattered roles.
A familiar count.Same company.
A different count.
01 — The question
For a century, the size of a company was the size of its payroll.
Companies have always been measured in people. Headcount went on the slide, the org chart filled the wall, and ambition was counted in desks. Accountants even gave it a unit: the full-time equivalent, one person’s working week. But look at what fills that week: drafting, researching, comparing, coding, testing, summarising, scheduling, following up. A growing share of it is work a model can now do, at any hour, many times over in parallel.
So the question changes. Not whether agents can do tasks, but whether a company can be composed of them. Not how many people it takes, but what only a person can do. zer0fte exists to find out, one product at a time.
The
spark.
Where do ideas come from when nobody is in the office?
One observation, many framings.
Branch an idea
64 paths · 1 choiceAgents widen the field. A person walks into it.
An illustration of divergence and choice, drawn fresh each time.
02 — The spark
Agents multiply the options. A person decides which one is worth a year.
Ideas still begin with someone noticing something: a workaround that shouldn’t exist, a question people keep asking, a tool that almost works. What changes is what happens next. One observation can become dozens of framings before lunch, each with a different customer, a different name, a different shape.
Agents are tireless at divergence. On their own, they are not good at caring which branch matters. That judgement, the mix of taste, timing and a sense of what deserves a year of someone’s life, is the part we keep human.
The agents
- Generate framings and names
- Sketch variations of one idea
- Find what already exists
The person
- Notices the problem
- Chooses what to pursue
- Says no to the rest
The
evidence.
What is true before anything is built?
Read until a pattern holds.
Keep reading
Fit 0.00A hunch. Evidence is the pattern that survives more reading. The outliers stay on the page.
An illustration, not a dataset.
03 — The evidence
Every claim arrives with a source and a date. Opinions wait their turn.
Plenty of products fail quietly, long before launch, from a belief nobody checked. Research was the step small teams skipped because it was slow. For a collective of agents it is the cheapest step there is. They can read the reviews, the forums, the pricing pages and the complaints, map who already solves the problem and how badly, and estimate how many people have it, all before anyone opens a design tool.
The discipline isn’t reading more. It’s writing down what was found, where and when, and keeping the evidence that disagrees. A verdict without its sources is just a mood.
The agents
- Read the market at volume
- Map competitors and their gaps
- Size demand from the ground up
The person
- Asks the question worth answering
- Weighs the verdict
- Decides whether to go on
The
plan.
What would make us stop?
Write the ending first.
Three paths
Written before step oneThe bar was missed. We agreed in advance what that would mean, so we stop.
Three paths, written down before the first one is taken.
04 — The plan
The reasons to stop are written before the reasons to start.
Plans usually describe success. Ours begin with failure: the specific numbers, by specific dates, that would tell us to walk away. Agents argue both sides, draft the scope, estimate the cost and name the risks a hopeful founder would rather not hear. Then a person signs, or doesn’t.
It sounds pessimistic. It’s the opposite. When stopping is cheap and agreed in advance, you can afford to start far more things, and the ones that survive have earned it.
The agents
- Argue for and against
- Draft scope and estimate cost
- Put numbers on the risks
The person
- Sets the bar
- Approves the spend
- Honours the stop
The
shape.
Can taste be delegated?
One screen is never just one screen.
Every state of one screen
4 states · 1 point of viewNothing here yet.
Start with one idea. We’ll take it from there.
Add an ideaThinking it through…
That didn’t work.
Nothing was lost. Try again, or come back later.
Try againDone.
Three changes made. Here’s what’s different.
See the changesOne screen, four states. Agents draft every one. A person decides how they should feel.
An illustrative screen.
05 — The shape
Taste stays human. The repetition doesn’t.
A product is never one screen. It is every screen in every state: empty before the first use, loading while it thinks, failing gracefully, succeeding quietly. Most of design is that unglamorous multiplication, and it is where agents excel: drafting the flows, the words, the edge cases, the contrast checks, the variations nobody has time for.
What they can’t supply is the point of view: what this thing should feel like, and who it is for. So a person sets the direction, and the collective fills the space that direction defines.
The agents
- Draft flows and every state
- Write the interface words
- Check contrast and accessibility
The person
- Sets the direction
- Decides how it should feel
- Cuts what doesn’t belong
The
build.
What changes when software writes software?
Watch the last line.
A build, start to finish
6 steps · 1 pause- 01specaccepted
- 02testswritten before the code
- 03buildpassing
- 04review2 issues found · 2 fixed
- 05stagingdeployed
- 06productionwaiting for a person
The last line never automates itself.
An illustrative log, not a transcript.
06 — The build
Speed is cheap now. Reliability is the craft.
Code used to be the bottleneck: a small team could only build what it had hands to type. That constraint is loosening fast. Agents write, test, review and refactor, and the distance between a description and a working build keeps shrinking.
But faster code is not finished software. Someone still decides what “working” means, sets what it may cost to run, and owns what happens after launch. So the build runs on three rules. Tests come before shipping. The stack is one a single person can maintain. And one step never automates itself: the move to production.
The agents
- Write and test the code
- Review each other’s work
- Keep it running, and cheap
The person
- Approves the architecture
- Decides when it ships
- Owns what happens next
Thereturn.
Small bets, scattered.
Growth that doesn’t wait for a hire.
What does growth look like without hiring?
An imagined installation
07 — The return
When costs follow usage instead of headcount, small bets become worth making.
In most companies, every new line of revenue eventually asks for a new line on the payroll. Growth and cost climb the same stairs. When agents do the work, the cost of trying something follows usage instead: you pay for what runs, not for who waits.
That changes the arithmetic of small bets. Products that could never justify a team can justify a test. Prices can be explored instead of guessed. Distribution, support and reporting can grow with demand rather than with a hiring plan. None of it is automatic. Every price, every dollar spent and every promise to a customer still passes a person. But the slope changes, and the slope is the business.
The agents
- Test pricing and packaging
- Run distribution and content
- Report the numbers, plainly
The person
- Sets the price
- Approves every dollar
- Answers to customers
The seam: you and the agents
With you 0
With the agents 0
Where does the person end and the agents begin?
Drag the seam. Watch what moves, and what never does.
08 — The seam
The gap between people and agents isn’t closed by removing people. It’s closed by designing the handoff.
The interesting part of a company built on agents isn’t the agents. It’s the seam: the exact place where a person hands work to a machine and takes it back. Put it in the wrong place and you get busywork for the human or unchecked risk from the machine. Put it in the right place and something new becomes possible: one person, with a founder’s judgement, working with the reach of a team.
That is what zero FTE means to us. Not zero people. Zero people doing work a machine could do, so that people can do the work only people can.
The next
company.
Three questions. One surface.
Two edges, you’d think. There’s only one.
Follow the blue line around.
Why is this possible now?
Models can now reason through long tasks, use tools, write and check code, and work side by side. No single model does it all. A collective does: different agents, often on different models, each checking another’s work. Each capability alone is a feature. Together they are the parts of an operating company. The cost of trying has fallen faster than the habit of hiring has changed.
What gets easier?
Starting. A business used to need a team before it had a single customer. Now the first version of many ideas can be researched, designed, built and put in front of people by one person and a collective of agents. More people can start. More ideas get a fair test. Failure gets cheaper, so there is room to learn from more of it.
What should never be automated?
Judgement about what’s worth doing. Responsibility for what was done. The trust a business runs on. The companies that last will be precise about where that line sits, and honest about it.
Start at
zero.
Every chapter. One question.
Read from the beginningWhere people meet agents.
Move the seamWhy this, why now.
The next company