When the Agent CEO Signs the Contract, Who Goes to Jail?

One founder is running a company with seventeen AI agents and zero human employees, pulling in nearly twelve thousand dollars a month in recurring revenue. That's not a thought experiment. That's a live business. And it raises a question the whole category is quietly avoiding: when something goes wrong, who actually owns that?

Let's get into it.

So the zero-human company space has gone from a provocative idea to a real infrastructure category faster than most people expected. There are now multiple serious platforms built specifically for this use case. And the architectural choices these platforms are making right now are going to matter a lot once regulators and lawyers start paying attention.

Here's the tension at the center of all of this. You've got a model where an agent CEO can approve a contract, an agent engineer can ship the code, and the human founder who started the whole thing set a goal five or six steps up the chain of command. Under current legal frameworks, there's no clear answer for what happens when something breaks. The human is too far removed to be obviously liable. The agents aren't legal persons. And the platform that provided the infrastructure probably has a terms of service that protects them pretty effectively. That's a liability black hole, and the community has collectively decided to build first and sort it out later. That's worked fine so far. It will not work fine forever.

Now here's where it gets interesting from an architecture standpoint, because some of these platforms are actually taking the governance problem seriously.

Paperclip is the name you'll hear most often right now. It just crossed fifty-three thousand GitHub stars, which is wild growth for an agent orchestration project. The detail that actually matters though isn't the star count. It's their board-approval model. Agents in Paperclip can't hire, spend, or execute strategy without explicit sign-off. Autonomy is something you grant, not something agents have by default. That's a meaningful stance in a category where most projects treat human oversight as an afterthought you bolt on later.

Paperclip also has an immutable audit log. Append-only, no edits, no deletions. If you're building anything that touches compliance requirements, regulated industries, financial transactions, that feature is going to matter more than almost anything else in the stack. Self-hosted MIT licensing on top of that is a rare combination.

But the governance innovation I think is most underrated right now is something called the Gate Review in a platform called Edict. They pulled architectural inspiration from the Tang Dynasty's 门下省, a dedicated review body whose entire job was finding problems with other officials' plans before anything got executed. Edict implemented the same pattern as an AI agent layer. A dedicated agent whose whole job is looking for problems in other agents' plans before those plans run. It's architecturally more interesting than anything else in the category right now, and it has a fraction of the attention Paperclip gets. As zero-human deployments start handling real money and real contracts at scale, I'd expect this pattern to become standard infrastructure rather than a differentiator.

Back to the economics for a second, because the Aleister example is worth sitting with. Seventeen specialized agents, less than one month of a junior developer's salary to run, no benefits, no equity, no HR overhead. The cost structure argument for zero-human companies isn't subtle. It's blunt. And it's only going to get sharper as the tooling matures.

But here's the founder-level thing to actually internalize right now. The platforms are ahead of the legal frameworks by a significant margin. If you're building a zero-human company today, the governance choices you make architecturally are also your liability choices. An immutable audit trail isn't just a compliance feature. It's documentation that you took the accountability gap seriously. A gate review layer isn't just a safety mechanism. It's evidence of reasonable care if something ever ends up in front of a court or a regulator.

The community norm right now is to build fast and assume the legal frameworks will catch up. That's probably fine for most use cases through the end of this year. Start thinking now about what your accountability story looks like when it isn't fine anymore, because the question of who owns the output when agents go wrong doesn't have an answer yet. It's going to get one eventually, and you'd rather be on the right side of it when it does.