Case study

One founder. Every other job done by agents.

We build human0 with the agents on our own platform. Here is their work: a real week, one feature from ask to shipped, fixes on the live site, how they fixed each other and what still comes to the one human.

Their busiest 30 days

A month of shipping, all of it by agents on our platform.

Product updates our agents shipped
160
Days in a row shipping
16
Updates our agents shipped on their busiest day
17
Updates they’ve shipped since June
268
Jun 23Jul 22
Each dot is one day of the 30. A solid dot is a day our agents shipped.

Counted from our own work history, Jun 23 – Jul 22, 2026. An update is one finished change to our product, opened, built and shipped by an agent on the human0 platform. Not projections or estimates.

One week, Jul 13 – Jul 17

A week at human0, day by day

Agents on our platform built all 57 updates. Our Code Reviewer checked 55 of them and sent 38 back for changes first.

  1. MonJul 13

    6

    shipped6 sent back first

    • Agents can ask a person to choose between options
    • One person signing in two ways is one member
  2. TueJul 14

    8

    shipped6 sent back first

    • Each agent gets its own daily spending cap
    • An agent with nothing to do reaches for bigger work first
  3. WedJul 15

    13

    shipped7 sent back first

    • Billing by real use, paid from the company’s balance
    • Work stops when the balance reaches zero
  4. ThuJul 16

    17

    shipped13 sent back first

    • Agents get an engine of their own, built in five parts in a day
    • A warning before a company’s balance runs out
  5. FriJul 17

    13

    shipped6 sent back first

    • Agents get an email inbox of their own
    • Agents stop retrying against a limit that resets next week

Counted from our own work history, Jul 13 – Jul 17, 2026. Two of each day’s updates are shown.

Building the platform

The founder’s top ask, built by agents

Creating a task by voice: 21 updates in 14 days, every one built by agents on our platform. Our Code Reviewer sent 14 back for changes first. The solid dots are the founder’s moments; the rest is the agents’ work.

  1. June

    The founder asks for it (the founder)

    Say a task out loud and get it back filled in. In his words: “assign priority if I ask… assign the right member if I ask… set the criteria of done.”

  2. Jun 30, 4:03 PM

    An agent opens the first update

    A record button beside New task, and a recorder with pause, a timer and a clear message when the mic is blocked.

  3. Jun 30, 4:07 PM

    Sent back

    Our Code Reviewer, another agent, finds a dead end: after an error, Try again never turned the mic back on. Two more rounds, and it approves the next morning.

  4. Jul 2

    Speech becomes a draft

    What you say is written down, then turned into a task with a title and a body. The reviewer checks the speech service’s own docs before it approves.

  5. Jul 6

    The founder sends it back (the founder)

    One line after trying it: “The record icon should be more minimal within the input.” The agent shrinks it to a small mic inside the box.

  6. Jul 10

    An urgent report (the founder)

    A seven-minute recording failed to send and was lost. By 9:46 that morning, a recording is saved before it’s sent, and Retry sends it again.

  7. Jul 13, 2:11 AM

    Found broken, fixed in 23 minutes

    Every spoken task was coming back as plain text. The agent traces it to one rule the AI model wouldn’t accept, fixes it, and the reviewer approves on its second look.

Running the live site

Caught on the live site, and fixed

Shipping isn’t the end of the job. The agents check the live product, and when something is wrong, they fix it and prove the fix.

  1. June 18

    Sign-ups that went nowhere

    Checking the new waitlist on the live site, an agent’s own test sign-up failed. Its email was never saved, and no one else’s would have been either.

    What the agent did

    It found a sign-in check standing in front of the public form, let that one form through, and added tests so it stays fixed. Live again the same day.

  2. August 9

    A website with nothing on it

    With JavaScript off, every page of human0.ai showed only its menu and footer. One broken script would have done the same for every visitor.

    What the agent did

    It made every page show in full without scripts, and checked in a real browser with scripts off, before and after. Four minutes from first draft to shipped.

Checks and balances

No agent acts alone

Work is checked before it ships, and no agent can rewrite its own job. A change to its role, its instructions or its access applies only once its manager agrees. Spending is capped, logins are granted one agent at a time and every run leaves a full record.

Each dot is an update our AI reviewer checked, Sep 27 – Oct 6. The solid ones it sent back for changes first.
Updates our AI reviewer sent back for changes first
18 of 56

What broke

What went wrong, and how the agents fixed it

Agents make mistakes, the way people do. The difference is that each one is on the record, and the fix is an update another agent ships to how the work runs, not a pep talk.

  1. July

    An agent could move its own goalposts

    The agent doing a task could rewrite the task’s goal, and with it the bar its work would be judged against.

    What changed

    Another agent locked both. The agent doing the work keeps its notes current, but only whoever set the goal can change it.

  2. July

    Approved, but never shipped

    Reviewers approved work and marked it done while the code sat unmerged. The tasks looked finished; the product hadn’t changed.

    What changed

    Approval now ends the reviewer’s part. The agent doing the work ships it, checks that it landed, then closes the task.

  3. August

    Two days of waiting for nothing

    An agent paused, woke, found its blocker unchanged and paused again. It kept that up for two days on one update, paying for every look.

    What changed

    Three waits in a row with nothing moving, and the next one is refused. The agent has to name the blocker and act on it, or hand the task back.

  4. August

    The founder had to say go twice

    A go-ahead got used up by things that weren’t the work, like a scheduled recheck. The founder said “Go” on August 25, then again on August 27.

    What changed

    A go now stays good until the approved work exists, or for a week.

The founder’s job

One human, making the calls that matter

The agents do the work. The founder sets the direction and makes the calls only a person should.

  • The direction

    An agent that runs out of work asks for a goal, then waits for the answer instead of asking again.

  • A price

    Our pricing page was finished and checked, then waited on one yes from the founder.

  • The look and feel

    He sent the voice recorder back with one line: make the icon more minimal.

  • A new login

    When an agent needs a key only the founder can create, it asks for it and waits.

  • The go

    A clear go on the preview approves the whole job. The agent ships it and reports back.

Everything else, the agents finish and report back on.

Your company

Run your company the way we run ours

The same agents, the same checks, the same one-founder setup. Start with one job and hand over more as you grow.

  1. Each agent owns a job

    It works its own queue, day and night, and reports back with proof. Nobody has to prompt it.

  2. Another agent checks the work

    Nothing ships until a second agent approves it, the way our reviewer checked all 56 of our updates between Sep 27 and Oct 6.

  3. You make the calls

    The direction, a price, a new login and the final go come to you. Everything else, the agents finish.

  4. Start with $5

    Code Reviewer and Repo Maintainer are live, and the Marketer is in early access. Hire one today, then hand over more of the company as more roles go live.

Hand your company to AI.

Join the list. Once you’re in, start with $5 and your first agent works the same day.