Zero-Human Companies Are Real, But They Can't Sell Anything Yet

Agents can build a full software product in 30 days. They cannot sell it to a single person.

That's the headline from a repo called make-money-30-day-experiment, published by Ithiel Labs, and it's the most honest zero-human company case study I've seen. The team spun up an AI agent stack, built 238 API endpoints, published 136 SEO pages, got listed on MCP registries. Thirty days in, revenue was exactly zero dollars. Not low. Zero.

And here's the thing, that failure is actually useful. Because it pinpoints exactly where the gap lives. Agents are genuinely good at construction now. They're still terrible at distribution. Bot detection walls, cold outreach filters, trust signals that only a human face can generate, these are the real moats. Not code. The insight from this repo is that 'agents can build anything' and 'agents can sell anything' are not the same claim, and we've been conflating them.

So what does the working version look like? YC W26 just handed us a dataset. Fourteen companies crossed a million dollars in annual recurring revenue with between one and four people on the team. Fourteen percent weekly revenue growth. These aren't vibes. Investors are actually putting a new metric into term sheets now called agentic task completion rate, which measures what percentage of a company's operational workload agents are handling end to end. That number is getting priced.

The pattern that emerges from W26 is that agent-native isn't a product category, it's a go-to-market strategy. These teams aren't building AI tools. They're running businesses where agents handle execution and humans handle judgment calls. Small teams, high leverage, and a very deliberate decision about which tasks need a human in the loop and which ones don't.

That judgment call also matters for a much more uncomfortable reason. There's a liability gap forming around zero-human companies and the courts are starting to find it.

You've probably seen the Air Canada chatbot case. Customer relied on what the chatbot said, airline tried to disclaim responsibility, court said no. An IBM refund agent case is pointing in a similar direction. The early signal here is that when an agent takes an action with real-world consequences, like approving a contract or issuing a refund, and something goes wrong, the founder of the company running that agent is the defendant. Not the framework maintainer. Not the model provider. You.

If you're running an agent stack right now without explicit liability architecture, meaning documented human approval thresholds, audit logs, defined scope limits on what agents can commit to, you're exposed. This isn't hypothetical anymore.

On the tooling side, OSS Insight mapped the entire zero-human company GitHub ecosystem and the architectural bets are diverging in interesting ways. Safety veto layers, cost-routing by task complexity, persistent agent memory, those are the three patterns separating frameworks that are actually production-ready from ones that are demos with a nice README. Paperclip is showing up with verticalized adapters. A marketplace called Clipmart is worth watching. The ecosystem is moving fast enough that the frameworks you picked six months ago might already be obsolete.

And then there's the Postiz founder Nevo David's marketing stack, which is the most concrete 'here's what I actually run' account in this space. Six layers of agent automation, two thousand new customers a month, one founder. The clever bit is using GitHub Issues as agent memory and a feedback loop, which is a genuinely smart workaround for the lack of persistent context across agent runs.

But here's the honest part at the end of his own writeup, and I think this is the most important sentence in all of this: it's not zero humans, it's humans as quality controllers instead of executors.

That's probably the more accurate framing for where we actually are. The zero-human company is a direction, not a destination yet. The teams winning right now have figured out which layer of their stack to fully automate and which layer still needs a human with judgment. The ones losing are either moving too slow on automation or moving too fast on liability.

The gap between building and selling is real. The liability gap is real. And the founders treating both seriously are the ones showing up in the YC data.