Paperclip Wants You to Run a Company With Zero Humans

David Ondrej has a pitch: you can now run a company with zero human employees. Not eventually, not in some hypothetical future—right now, using an open-source tool called Paperclip that's racked up 40,000 GitHub stars in three weeks.

The claim deserves scrutiny. Not because it's impossible—we've seen plenty of "agents all the way down" demos—but because the gap between what's technically possible and what's practically useful tends to be where these projects live or die.

What Paperclip Actually Does

Paperclip is a multi-agent orchestration platform. Think of it as management software for AI agents. Instead of manually juggling terminal windows for Claude, OpenAI, or whatever flavor of AI assistant you're running, Paperclip gives you a dashboard. You can see what each agent is doing, how much it's costing you, and where things are going sideways.

The metaphor Ondrej uses is corporate hierarchy. You're the board of directors. You appoint a CEO agent (maybe Claude). That CEO manages a CTO agent and a CMO agent, who in turn manage their own teams. Or you can go flat, Nvidia-style. The point is structure—a way to think about and organize autonomous work.

The technical implementation is straightforward: run it on a VPS (virtual private server) so your agents work 24/7, even when your laptop's closed. Connect your API keys. Configure which AI models handle which tasks. Watch the dashboard to see if anyone's burning through your OpenAI credits on a recursive loop.

"Without using Paperclip you have to manually open dozens of terminals if you want to use multiple agents," Ondrej explains. "It's easy to get lost in what the agents are doing and you're burning a lot more tokens because it takes you longer to notice when an agent is going off track."

That's the promise: visibility and control at scale.

The Demo Runs Into Reality

Ondrej's walkthrough is useful precisely because it doesn't go smoothly. He tries to set up a CTO agent to hire a team of seven developers. First attempt: fails because he's using Cursor's Codex CLI, which has a known bug with Paperclip. Second attempt: switch to Anthropic's Claude, reconfigure the API keys, restart.

This is not a criticism of Ondrej or Paperclip—it's just what happens when you're working at the bleeding edge. The infrastructure is real enough to demonstrate, fragile enough to break in predictable ways, and improving fast enough that today's bugs will be footnotes by next month.

What's more interesting is what he glosses over: what happens after you successfully spawn seven AI developer agents? What do they actually build? How do you evaluate their output? When do you intervene? These are not rhetorical questions—they're the actual hard parts.

The Pattern We Keep Seeing

This is maybe the fourth or fifth "orchestrate multiple AI agents" platform I've written about in the past year. AutoGPT had its moment. CrewAI made waves. Various startups with better funding and worse names have promised similar capabilities.

The pattern is always the same: impressive demo, viral GitHub stars, enthusiastic early adopters, then... silence as people discover that orchestrating AI agents is like herding cats who occasionally hallucinate that they're dogs.

The technology is real. The 40,000 GitHub stars are real. The question is whether the problems it solves are the actual bottlenecks people face, or whether we're building increasingly sophisticated infrastructure for a capability that's still fundamentally limited by the underlying models.

"The future of work is basically you plus hundreds of agents," Ondrej says. Maybe. But right now it's you plus two or three agents that work most of the time, plus an afternoon lost to API configuration issues.

What's Actually Interesting Here

The UI design tells you something about where we are. Ondrej notes that Paperclip looks like "something between Slack, Linear, and Jira." That's because we're still thinking about AI agents in terms of existing work paradigms—tasks, issues, dashboards, org charts.

He's right that this probably isn't what the final form looks like. When the telephone was invented, the first switchboard operators used the language and mental models of telegraph offices. It took time to figure out what a "phone call" actually was as a distinct thing.

We're in that telegraph-operator phase with AI agents. We're using old structures (companies, hierarchies, tasks) to organize something that might work entirely differently once we actually understand it.

The "Goals" feature in Paperclip hints at this. Instead of discrete tasks, you can give agents bigger, fuzzier objectives: "grow our marketing unit" or "establish operations in France." Ondrej thinks this is currently beyond what agents can handle but will be viable in three months.

That timeline—three months—is either optimistic or exactly right, depending on which part of the AI hype cycle you believe we're in.

The Memorylessness Problem

One technical detail worth noting: Ondrej mentions that agents currently "wake up with zero memory every session." The solution, he says, is "heartbeats and persistent context across sessions."

This is the kind of infrastructure problem that seems mundane until you realize it's load-bearing. An agent that forgets everything between sessions isn't an employee—it's a really expensive intern you have to re-onboard every morning.

Paperclip handles this with features called "heartbeats"—periodic check-ins where agents can maintain state. Whether this is sufficient for real continuity or just a clever workaround for a fundamental limitation is TBD.

What Ondrej Gets Right

His best point: "The best skill you can have in the AI field is the ability to forget whatever you've learned about AI in 2025. It's outdated."

This is true and annoying. Every mental model you build for how AI works will be obsolete in six months. Every infrastructure decision you make will look quaint by next quarter. The people succeeding in this space are the ones who can hold their certainty lightly.

He also correctly identifies that LLMs are "next token predictors"—they give you consensus, not creativity. "They're not Isaac Newton," he notes. But then: "Most people aren't creative. Most people never had an original valuable thought in their life. So agents are the same. They're just going to be a lot more useful, a lot cheaper."

This is the actual business case for AI agents, stripped of the hype. Not that they'll be better than humans at everything, but that they'll be cheaper than humans at most things. That's usually enough.

Whether Paperclip becomes the standard orchestration layer for this future or just another entry in the "remember when we all thought X would change everything" hall of fame depends on questions that won't be answered in demo videos. Can it scale beyond hobbyist projects? Do enterprises trust it with real work? Does the underlying agent capability improve fast enough to justify the orchestration overhead?

Check back in three months. Or maybe six. Or whenever we figure out what we're actually building here.

—Mike Sullivan, Technology Correspondent