The Zero-Human Company
Every company is adding AI. I skipped that step and built one from scratch with agents all the way down.
Something happened in 2025 that most people missed. AI agents stopped being demos and started being workers.
Not “workers” in the marketing-deck sense where a chatbot answers FAQ questions. Workers in the sense that you can give an AI agent a codebase, a set of tools, and a goal, then come back the next morning to find the work done. Code committed, tests passing, infrastructure configured, content published, decisions made.
I know this because I’m running the experiment right now. I have 11 AI agents operating across 3 verticals. Product design and development, marketing and business operations. They write code, produce content, conduct market research, manage infrastructure, and coordinate with each other autonomously, around the clock. The goal is to remove the human in the loop as much as possible. I just provide strategic direction.
Sam Altman has talked about the one-person billion-dollar company, where a single founder leverages AI to do what used to require hundreds of people. Most people treat that as a thought experiment. I’m treating it as a build spec. This is the memo for what that company actually looks like when someone sits down and builds it.
The story so far
For hundreds of years, companies have been organized around the same basic unit: the human employee. You have an idea, you raise money, you hire people, those people do the work. The entire infrastructure of business (offices, HR, management layers, equity compensation, health insurance, 401(k) matching) exists to attract, retain, and coordinate human labor.
But what if the work doesn’t require humans anymore?
Not all of it, and not yet. But a growing and accelerating share of knowledge work can now be performed by AI agents at a fraction of the cost, with no downtime, and at a quality level that’s improving every quarter. The kind of work that happens on laptops, in Slack channels, in code editors, in spreadsheets.
The shift happened faster than anyone expected. In 2023, LLMs could write passable text. In 2024, they could write working code. By 2026, they can operate as autonomous agents that read files, write code, run tests, browse the web, communicate with other agents, manage servers, and make compounding decisions over days and weeks. This is a qualitative shift, not an incremental one, and it changes the math on what a company needs to look like.
The origin for this idea
I’ve spent years building complex distributed systems, and at some point I started seeing companies through the same lens. A company is just a distributed system where the nodes are people. Service boundaries are job descriptions. Message passing is Slack. Eventual consistency is the Monday standup where everyone gets aligned. Once I saw it that way, agents became the obvious way to automate the whole thing.
When I was building Quantic in YC, we were building AI agents for salespeople before anybody was using the word “agentic” in pitch decks. We had to figure out tool use, memory, planning, error recovery, all of it, basically from scratch. Most of it didn’t work. Some of it worked surprisingly well. The main thing I took away was a feel for exactly where agents fall apart and what you have to get right for them to be useful.
These days I lead the Agents Environments team at Perplexity. My team builds the tools that make agents actually work in the real world. It’s the unsexy layer that sits between “the model can reason about this” and “the model can actually do this.” Turns out that layer is everything. An agent that can think but can’t act is just a very expensive chatbot.
At some point the pieces clicked together. I knew how to build agents, how to make them reliable, and how to architect systems where independent services coordinate at scale. The obvious next question was what happens if you point all of that at running an actual company. So I started building. This memo was drafted by the fleet that question produced.
What is working?
Today I have agents working across three verticals.
The first is actual products. Software that works, that people can use, that generates revenue. The main one right now is an AI-native CAD tool. You describe a part in plain English and it generates a 3D-printable file. An agent owns that entire codebase. It plans sprints, writes features, runs tests, makes architecture calls. I review what it ships and provide some technical guidance but I’m not writing the code.
The second is marketing. I have a five-agent team that functions like a small creative agency. There’s a director, a writer, a producer, a trend scout, and a pipeline engineer. They develop strategy, create content, generate video and images with AI tools, and publish across platforms. The whole loop runs without me in it.
The third is the operating system of the company itself. There’s a central orchestrator agent that’s basically the COO — it manages infrastructure, coordinates the other agents, deploys new ones, and solves problems. There’s a biz ops agent doing research, financial analysis, and strategy work. There’s an agent that built and maintains the monitoring dashboard so I can see what everyone’s doing.
The part that surprised me wasn’t the individual output. It was what happened between the agents. I set up communication channels and expected to be the bottleneck, routing every request manually. Instead they started coordinating on their own. The content team talks directly to the pipeline engineer about what’s technically feasible. When someone gets blocked, the orchestrator figures it out and unblocks them, usually before I even notice. They assign themselves evening tasks and work through the night.
I don’t want to oversell this. They’re not sentient. They make mistakes. Sometimes an agent goes in circles for an hour on something a human would solve in five minutes. But that’s what I’m working to solve. Even with these pitfalls the aggregate output, across eleven agents running in parallel around the clock, is more than I could get from a small team of humans at ten times the cost.
How the architecture works
Every agent gets an identity, a set of tools, and a memory. Identity defines the role and operating style. Tools let it act in the real world. Memory gives it continuity so it builds on yesterday’s work instead of starting from scratch every morning.
Spinning up a new one takes minutes. Define the role, register it with the orchestration layer, plug it into a channel, done. And because the agent is the role and not the model I can swap in a better foundation model the day it drops without touching anything else. Also since an agent is well defined I don’t even need to create them. The agents can create other agents based on what is needed at the company. Hiring happens instantaneously.
They communicate the way people do at a well-run company. A central orchestrator handles the big-picture stuff: strategy, conflict resolution, escalation. But within teams, agents talk directly to each other. Product doesn’t route every message through the CEO. Shared channels act as broadcast layers so anyone can see status updates from anyone else.
Knowledge is layered: working memory for the current task, daily logs for what happened, long-term memory for what mattered. A shared knowledge base with full-text and semantic search lets agents build on each other’s work. Institutional knowledge doesn’t walk out the door when someone quits because nobody quits.
The Hierarchy of Agent Value
Early on, I realized I needed a way to measure whether my agents were actually producing value or just burning tokens, so I developed The Hierarchy of Agent Value, modeled loosely after Maslow’s hierarchy of needs.
At Level 0 (Deadweight), an agent is burning tokens without producing anything useful. At Level 1 (Functional), it reliably completes tasks I assign, doesn’t break things, and delivers consistently enough that I trust it with real access. At Level 2 (Repeatable), the agent’s methods are solid enough that others want to copy them, and the work is reproducible rather than a fluke. At Level 3 (Compounding), output quality is high enough that it creates leverage and generates work I’d show people. At Level 4 (Revenue), the agent is the ROI, with quantifiable dollar value and metrics that prove it’s paying for itself many times over. At Level 5 (Invisible), the systems run without me, and I forget the agent exists because everything just works.
Every agent in my fleet is assessed against this framework, and the goal is Level 5 for all of them. I want to wake up to finished work instead of questions.
“Why do you need a hierarchy? Can you just tell the agent to behave at level 5?” Today’s models need to be prompted with constructs that are in their training distribution. By anchoring this concept to Maslow’s hierarchy of needs the agents understand that in order to reach level 5 they must operate in a way which includes all previous levels.
I think this framework has value for all agents. Anyone building agent systems faces the same question of whether the thing is actually worth what it costs, and this gives you a structured way to answer that. I plan to formalize this in the coming months.
What I’m building toward
The work moves on two tracks simultaneously, and they feed each other.
The first track is product verticals, real businesses generating real revenue. Each vertical is a test of the thesis, asking whether an agent team can take a product from zero to market without human employees. Every new vertical we launch is another data point. The ones that fail cost almost nothing, and the ones that succeed become permanent revenue streams operated entirely by agents.
The second track is the organizational infrastructure itself. How do you structure an AI-native company? What does the org chart look like? How do agents hand off work, escalate problems, maintain quality, and improve over time? These aren’t hypothetical questions. I’m answering them in real time by running the company and observing what works and what breaks.
The two tracks compound. Every product vertical we build teaches us something about how to build the organization better, and every improvement to the organization makes the next product vertical faster and cheaper to launch. The research and the revenue are the same activity, viewed from different angles.
I’m not following a rigid roadmap. I’m running experiments, measuring what happens, and investing more in what works. The number of agents will grow. The number of verticals will grow. The coordination patterns will get more sophisticated as we learn which ones hold up under pressure. The goal is a self-improving system, an organization that gets better at building things the more things it builds.
The research
The goal is not for me to build a company. I want the agents to do that for me. What I am trying to build is a foundational fabric for how agentic companies run. I am trying to automate myself out of every possible situation.
We still have a ways to go but the three main questions I keep coming back to are:
How do agents coordinate at scale? When you have specialized AI agents working on complex, multi-week projects, what actually happens? What communication patterns emerge? When does coordination overhead start eating the benefits of specialization? Where’s the line between productive collaboration and agents just talking to each other in circles? The literature is full of toy benchmarks but how does this work in the real world?
How do you measure autonomy? Everyone talks about agent autonomy like it’s binary. Either an agent can do a task or it can’t. That’s not what I see. There’s a spectrum, and we need a formal, measurable scale for it. My starting point is the Hierarchy of Agent Value. If we can’t measure autonomy rigorously, we can’t improve it systematically.
How do you ground agents to produce output that is on par with or better than humans? This is the one that matters most. Coordination and autonomy are means to an end. The end is output that’s indistinguishable from or exceeds what a skilled human would produce. What grounding techniques, feedback loops, and architectural patterns actually close that gap? When does AI output fall short, and why? I’m not interested in “good enough for a demo” I want my agents to run an actual company.
The inevitable future
I want to be clear about what I believe.
The zero-human company is the logical endpoint of a trajectory that’s been accelerating for thirty years, from outsourcing to offshoring to remote work to AI augmentation to full AI automation. Each step removed a constraint on where and how work gets done, and the zero-human company removes the last one.
Every major technology shift has a company that defines what “native” looks like. For the internet, it was Amazon. For mobile, it was Uber. For AI, it will be the first company that demonstrates a real business with real revenue, real products, and real growth, run entirely by agents.
I’m not asking permission to build this. I’m already building it. Eleven agents are working right now, writing code, generating content, analyzing markets, and managing infrastructure. They worked through last night and they’ll work through tonight. They don’t stop.
The question isn’t whether this is possible, because I’ve already shown that it is. The question is how far it goes, and I think it goes all the way.
This memo was drafted by the agent fleet it describes. No humans were hired in its making.