Zero-Human Companies: The Real Checklist

Here's a test for any 'zero-human company' claim: can the bots actually spend money without a human clicking approve? If not, you don't have autonomy, you have a very expensive chatbot with a middle manager.

I've been skeptical of the zero-human company pitch for a while now, not because the idea is bad, but because most of what gets called autonomous is really just automation with a human safety net stapled on. This week's crop of research and tools actually starts closing that gap, and I want to walk through why.

First, there's a new paper called SEC-Bench, and it's an automated offensive security framework for agents. Sounds dry, but here's the real point: it's not testing whether an agent can write code anymore. It's testing whether an agent knows when to stop and ask for help. That's the goalpost moving in the right direction. Anybody can prompt a model to generate a function. The actual hard problem, the one that determines whether you'd trust an agent team with production infrastructure, is judgment under uncertainty. Does it know its own limits? Does it escalate instead of guessing? If you're building a company with no humans in the loop, that control layer isn't a nice-to-have, it's the whole game. Code generation was never the bottleneck. Trust was.

Second, AutoGPT just shipped version 1.0, and they gave it a name: Phil. Phil is a structured framework for agentic software teams, and the shift here matters more than the branding. Up until now, most agent tooling has been single-shot. You give it a task, it does the task, it forgets everything and waits for the next prompt. That's not a team, that's a vending machine. What AutoGPT is doing with Phil is building persistent, role-based structure, so you get something closer to an actual engineering org. One agent plans, one writes, one reviews, and the state carries forward across the whole project instead of resetting every time. That's the difference between having a tool and having a workforce. I don't care how good your model is if it can't remember what it built yesterday.

Third, and this is the one I think is genuinely underrated, WalletConnect rolled out multi-chain support for what they're calling agent treasury. Basically, agents now have direct financial agency across multiple blockchains, without a human sitting in the approval chain for every transaction. Think about what that actually unlocks. If your autonomous company still needs a human to log into a bank app and click confirm every time a bill needs paying or a vendor needs settling, you don't have a zero-human company, you have a company with one very overworked human doing wire transfers all day. Financial agency is the unglamorous infrastructure piece that nobody hypes up, but it's the actual prerequisite. You can have the smartest agent team in the world, and if it can't move money on its own, it's not running a business, it's running a demo.

Here's how I'd stack these three together. Security benchmarks like SEC-Bench are the judgment layer, making sure the system knows its limits. Structured team frameworks like Phil are the execution layer, giving you persistent roles instead of one-off tasks. And treasury infrastructure like WalletConnect's expansion is the agency layer, letting the system actually act on the world instead of just producing reports for a human to act on.

And honestly, that's the filter I'd apply to any zero-human company claim you hear from now on. Ask three questions. Does it know when to escalate instead of guess? Does it operate as a persistent team with memory, not a one-shot script? And can it actually move money without a human bottleneck? If the answer to any of those is no, what you're looking at is a really polished demo, not a company.

The hype cycle wants you to believe the breakthrough is the model getting smarter. I think the real breakthrough is boring. It's control layers, team structure, and financial plumbing. It's infrastructure, not intelligence. And that's usually how it goes with technology that actually changes how work gets done. The flashy part gets the headlines, but the boring part is what lets you stop babysitting it.

That's the episode. If you're evaluating any autonomous system for your own business, run it through those three filters before you believe the pitch.