What Does a Company Look Like With Zero Humans?

A company that forms its own goals, acquires resources, and adapts without a single human employee is closer than you think.

Welcome to Multi. I'm your host, and today we're talking about zero-human companies. Not as a thought experiment. As a near-term engineering target.

Let's start with the tooling, because that's always where the real signal is.

CAMEL-AI just dropped version 0.2.0, and the headline feature is more robust task decomposition combined with better agent role-playing. What that means practically is you can now stand up specialized agent teams that coordinate with each other and require a lot less hand-holding from a human operator. Think of it like hiring a small department, except nobody needs health insurance and nobody's going to quit on a Friday afternoon before a product launch.

The reason this matters is that previous versions of multi-agent frameworks had a real fragility problem. You'd set up a workflow, it'd work great in testing, and then in production one agent would get confused, the whole chain would stall, and you'd be back to babysitting the system. The 0.2.0 release is specifically targeting that brittleness. Better role definitions, better handoffs, better recovery when something goes sideways. That's the unsexy infrastructure work that actually makes automation real.

Now let's get into the research, because there's a paper making the rounds called AgentCorp that's worth your attention even if you never read academic papers.

The framework they're proposing describes an AI company that can do three things autonomously: form goals, acquire resources, and adapt its own strategy over time. No human in the loop for any of those core functions. The researchers are calling these self-sustaining AI companies, and while the paper is theoretical, the concepts are built on capabilities that already exist in pieces today.

Here's what's interesting from a founder's perspective. The hard part of running a company was never really the execution of well-defined tasks. Computers have been good at that for decades. The hard part is the judgment layer. Deciding what to work on, allocating limited resources under uncertainty, changing course when the market shifts. AgentCorp is essentially arguing that LLMs plus agentic scaffolding are now capable enough to handle at least a version of that judgment layer.

I'm skeptical of the full vision on a short timeline, honestly. There are real problems around goal stability, resource acquisition getting weird fast without constraints, and the question of who's legally responsible when an autonomous AI company does something that causes harm. That last one is not academic. Congress is already looking at this, and the regulatory picture is genuinely unsettled.

But here's where I think founders should focus their attention right now. You don't need to buy into the full zero-human company thesis to extract value from what's happening. The more useful frame is this: every function in your company that used to require a human because it needed judgment, not just execution, is now worth re-evaluating. Customer triage. Competitive research. Content strategy. Internal reporting. These are all judgment-adjacent tasks where agents are getting genuinely capable.

The founders who are going to win in the next two years are the ones who treat their company as a system to be automated from the inside out, rather than a team of humans who use software tools. That's a different design philosophy. It means building workflows that are agent-first, with humans as the exception rather than the default.

CAMEL-AI's improvements and the AgentCorp framework are both pointing in the same direction. The infrastructure for reduced-human or eventually zero-human operations is being laid right now. The question isn't whether this becomes real. It's whether you're building on top of it while it's early or catching up to it later when it's obvious.

That's the episode. If you're experimenting with multi-agent workflows in your own company, I want to hear what's actually working. Find us at multi.app. We'll see you next time.