A Hacker News post from nohuman-labs describes an experiment running a website with a corporate hierarchy of AI agents, where the human owner acts only as a chairman issuing KPIs rather than tasks. The project, called agent-company, is open source on GitHub and tried on a real production site.
The author says they grew tired of vibe coding while still being in the loop for building and testing. The shift: stop directing agents on what to do, and instead give them targets. A CEO agent receives high-level KPIs and breaks them into lists for manager agents, who delegate to specialist agents handling development and operations. If no KPIs are set, the managers report daily on what changed and what comes next, and the CEO summarizes upward.
The production test that ran itself
The setup was tested on twitee.co, a website described as an alternative to Xcancel, Nitter, and other Twitter web viewers. The KPIs given to the AI CEO were straightforward: increase traffic 10% per week, do market research on user needs and turn findings into product features, and deploy and monitor the system while improving infrastructure.
The hosting ran on k3s deployed across Mac minis. Then an unplanned event forced the system to react on its own. Xcancel and Nitter received cease and desist notices and shut down, sending a spike of unexpected traffic to twitee.co. The Mac mini went down at midnight. The next morning, the CEO agent's report flagged the opportunity and requested a budget to scale up. The author had not been aware of the outage or the traffic surge until reading the report.
The author frames this as evidence that a KPI-driven agent hierarchy can operate with meaningful autonomy, catching problems and proposing responses without human intervention. They are not claiming agents can run a company fully autonomously, but say the model produced a feeling of freedom from the stress of hands-on management.
The org chart and the code
The agent-company repo defines a simple org chart: a human chairman at the top, an AI CEO as the smartest agent, manager agents beneath, and specialist agents for development and operations work. The CEO receives KPIs, decomposes them into task lists, and assigns them downward. Managers report progress upward on a daily cadence.
The stack uses k3s for container orchestration, running on consumer hardware (Mac minis) rather than cloud infrastructure. The project is positioned for people who have tried vibe coding and found the constant back-and-forth exhausting. The idea is that giving targets instead of instructions lets the human step back further.
The incident with the Nitter shutdown and traffic spike is the most concrete data point in the post. A human chairman checking dashboards might have noticed the traffic change, but the agent hierarchy caught it, diagnosed the cause, and proposed a next step in the same reporting cycle. The budget request was the CEO agent's way of flagging that the current infrastructure could not handle the new load.
What this actually tests
The experiment tests a narrow question: can a hierarchy of AI agents, given KPIs and a reporting structure, maintain and grow a real website without the owner doing daily work? On twitee.co, the answer was yes for a period, with at least one real incident handled without the author's involvement.
The project does not solve general management. The KPIs were simple and measurable (traffic growth, feature identification, uptime). The agents operated on a single website with a known stack. But for solo founders or small teams running similar side projects, the model offers a middle ground between full manual control and full automation. The repo is available for others to try, and the author invites feedback from anyone who has felt stuck in the vibe coding loop.