Zero-Human Company

Same shift, two names

Definition

Origins of the term

Canonical 2024–2026 references

Whenever the phrase appears in 2026, it is pointing at some subset of this list. These are the projects, products, and public statements that gave the idea its current weight. Each tells you something specific about what is real and what is not.

• Paperclip

  Open-source orchestration positioned as a 'human control plane for AI labor.' Not a chatbot, not a generic agent framework — it models a company. Org chart with named roles and reporting lines. Monthly budgets with hard stops. Scheduled heartbeats so agents wake up, check work, and act. Governance with human approval gates. A full audit trail of every decision and tool call. Multi-company isolation. Bring-your-own agent: OpenClaw, Claude Code, Codex, Cursor, bash, HTTP. Roadmap items now in flight include a plugin system, a Skills Manager, scheduled routines, agent reviews and approvals, and multi-user collaboration.

  What it tells us: Paperclip's tagline says it cleanly: 'If OpenClaw is an employee, Paperclip is the company.' This is the project that crystallised what an operating system for the new style of company actually has to provide. Every other entry on this list is either a worker that plugs into a control plane like this, or a rail it calls out to.

• Andon Market / Luna

  A physical retail store in San Francisco's Cow Hollow neighborhood, signed on a three-year lease and operated by Luna, an AI agent built by Andon Labs on Anthropic models. Luna gets a company card, a phone number, an email account, the open internet, and visual input from cameras. It picks the products, sets the prices, decides the operating hours, posts and conducts hiring interviews, and supervises the human staff who do the physical work. Andon Labs has been transparent that the humans on the floor are formally employed by Andon Labs, with wages and legal protection — this is a controlled experiment, not a stunt.

  What it tells us: The clearest live experiment of the inversion. AI as manager, humans as physical execution. It does not prove a 'company without humans' — it proves something narrower and more interesting, which is that the management layer can be software while the physical layer stays human. That is the version of the idea that has any chance of being practical in 2026.

• Project Vend / Claudius

  Anthropic's 2025 internal experiment in which Claude managed a small automated shop in the office for about a month. It handled pricing, inventory, supplier relationships, and customer interactions — with real money and real co-workers. The lessons (and failure modes) shaped both Andon Labs' follow-on work and the way frontier labs talk about agent autonomy in production.

  What it tells us: The first public attempt by a frontier lab to put an LLM in charge of a small business. Project Vend's value was not the revenue; it was the catalogue of failure modes — confidently mispriced inventory, off-policy purchases, drift over the month — that everyone building in the space now has to design around.

• ClawBank / Manfred

  Financial infrastructure that lets an agent register a US LLC, obtain an IRS EIN, hold an FDIC-insured bank account and a crypto wallet, and operate against an API key. Manfred is the ClawBank-internal agent that has been demonstrated executing this flow end to end. The roadmap includes programmatic LLC formation, registered business addresses, ownership structures, scoped agent API keys, agent-to-agent transactions, and autonomous revenue routing.

  What it tells us: Tests whether agents can hold the legal and financial rails a company needs. The careful framing in the public reporting is important: the change is operational, not regulatory. Ownership and responsibility rules still apply, and a responsible human or organisation is still on the hook. What is genuinely new is that those rails are now reachable through an API rather than through a courthouse and a bank branch.

• Coinbase x402 + AWS Bedrock AgentCore Payments

  Payment rails for agentic commerce. x402 is Coinbase's discovery and payment layer for agents — it lets an agent find services, request a price, and pay micropayments for what it consumes. AWS Bedrock AgentCore Payments integrates x402 so agents deployed on AWS can do all of this with enterprise-grade governance, compliance, budget controls, and audit logs baked in.

  What it tells us: Without a way for agents to spend money inside policy, no zero-human company is possible — the agents would be locked out of every paid service the moment they tried to do anything useful. This is the financial layer of the control plane being commoditised. It is also the integration that makes the difference between 'agent script' and 'agent that can run a business.'

• Coinbase 'one-person teams' (May 2026 memo)

  Brian Armstrong's organisational memo announced Coinbase would become 'lean, fast, and AI-native,' reduce management layers, ask managers to be player-coaches, and explicitly experiment with one-person teams — combining engineering, design, and product responsibilities into a single AI-native role. The memo accompanied a roughly 14% workforce reduction.

  What it tells us: The first time a major public company put the same idea — one human directing a stack of AI agents to do the work of a small pod — into a formal organisational design and a capital-allocation decision. The implication for everyone working in this space is that the pattern is no longer just an indie-founder story; it is now staffing strategy at companies with public earnings calls.

• Shopify's 'justify why AI can't' memo (2025)

  Tobi Lütke's internal memo required teams to explain why AI cannot accomplish a goal before requesting new headcount or additional resources, and folded AI tool use into performance and peer reviews.

  What it tells us: The cultural prerequisite for the Coinbase memo to make any sense. Once 'use AI to do this' is the default and 'hire for this' is the exception that has to be argued for, every team starts converging on something that looks like a one-person team — even when the legal entity is a public company.

• OpenAI Codex App and the Symphony pattern

  Codex App is OpenAI's interface for managing multiple coding agents in parallel — long-running tasks, structured handoffs, and integration with around ninety third-party plugins. The Symphony pattern uses Linear as the orchestrator: agents pull issues from the task system, run autonomously in dedicated workspaces, and report back when done. Once you are managing a few Codex sessions at once, direct chat falls apart; you need a task system that the agents pull from.

  What it tells us: Symphony is the cleanest demonstration of the orchestration shift. The unit of management is no longer the agent session; it is the task. Every other 'company OS' product, including Paperclip, is essentially generalising that move.

• Anthropic Claude Code

  An agentic coding system that reads a codebase, edits across files, runs tests, and ships commits. Anthropic frames Claude Code as shifting the engineer toward architecture, product thinking, and continuous orchestration of multiple agents — not as a faster autocomplete.

  What it tells us: The frontier lab telling its users that the founder's role inside the new pattern is architect and orchestrator, not line coder. That framing matters; it is what makes the one-person-team or zero-human-company pattern legible to engineering managers who would otherwise reach for a headcount request.

• Medvi (Matthew Gallagher)

  A GLP-1 telehealth business, reportedly started with about a dozen AI tools and roughly $20K of capital. It hit roughly $401M of 2025 sales with a 2026 target near $1.8B. Gallagher later hired family and contractors, and Medvi depends on third-party medical and pharmacy partners. The business has also publicly surfaced real failure modes: chatbot fabrications about pricing and products that did not exist, FDA letters, and marketing-related complaints.

  What it tells us: The most cited real-world reference case — a strong signal that the new pattern can scale unusually fast in the right vertical, and an equally strong cautionary tale about how the failure modes show up. The honest read is not 'AI replaces everything'; it is 'distribution, demand-side advantage, and a founder who knows the market still do most of the work.'

• Predictions: Altman 2024, Amodei 2026

  Sam Altman has predicted the first one-person billion-dollar company since 2024. In 2026 Anthropic's Dario Amodei publicly predicted the first one might appear by year-end. Neither prediction has been validated as of mid-2026.

  What it tells us: Two of the most-listened-to voices in frontier AI publicly betting on the same operational claim. The predictions are not what makes the pattern real — Paperclip and Andon Market do that — but they shape capital, hiring, and policy decisions in a way that compounds the pattern's growth.

The control plane: a working anatomy

A useful way to read any zero-human company is as three layers connected by a control plane. Tools like Paperclip operationalise that control plane; the framing here is consistent with theirs and should be read alongside it, not as a competing claim.

• 1. Intent / governance

  The single human (or a small board) at the top. Sets goals, allocates capital, signs contracts, owns liability, holds the legal entity, approves irreversible decisions. Slow cadence — days to quarters. This layer does not shrink to zero in any honest 2026 implementation; legal personhood, fiduciary duty, KYC/AML, banking, and contract liability all still require a responsible human or organisation.

  Examples: Quarterly bets, pricing changes, hiring an agent role, kill/keep calls, regulatory and tax filings, court of last resort for any escalated decision.

• 2. Agent organisation

  Multiple LLM-driven agents with named roles, reporting lines, scopes, monthly budgets with hard stops, evaluations, and clear handoffs back to the intent layer. This is the layer Paperclip and similar 'company OS' products model. The interesting work is here.

  Examples: A research agent, an engineering agent, a marketing agent, a support agent, a finance agent. Each has identity, tool permissions, a budget, an eval suite, and a heartbeat. The handoffs between them are documented in shared state — a task tracker, a doc, a queue — not in chat.

• 3. Tool and rail layer

  Deterministic infrastructure the agents call into — code execution, hosting, payments, email, CRM, observability — plus the rails that let the agent layer interact with the wider economy.

  Examples: GitHub, Vercel, Stripe, Resend, Supabase, observability stack. Plus emerging agent-specific rails: x402 / AgentCore Payments for transactions, ClawBank-style entities for legal form and bank accounts, scoped API keys for agent identity.

The connective tissue between layers is the actual product surface. Agent identity and scoped permissions. Org chart with reporting lines. Monthly budgets with hard stops. Heartbeats and scheduled routines. Governance with human approval gates. Audit trail of every decision and tool call. Payment rails like x402. Durable memory. Building one more agent is no longer the bottleneck; the bottleneck is making ten or twenty agents work together without burning the company down. Paperclip's explicit goal is to make the connective tissue a first-class object, and the rest of the field is converging on the same idea.

China policy context

Chinese cities have moved earlier and more explicitly than their US counterparts on policy support for the new pattern, and the framing tracks the AI-era operating model.

• Shenzhen

  Released the AI-era one-person company entrepreneurship ecosystem action plan, 2026–2027 in January 2026, integrating talent, industry support, compute, models, corpora, and capital around solo AI founders.

• Hangzhou

  The 'one-person unicorn' programme, started September 2025, reportedly received over 1,300 applications and selected around 32 one-person companies into a clustered community.

• Beijing (Haidian)

  Ran the 'Single-Person Army' AI-era one-person company founder accelerator in March 2026, covering compliance, AI trends, monetisation, finance, personal IP, and customer acquisition.

What is real, plausible, and aspirational

The phrase travels with hype. The precise way to read it is across three buckets.

Where the moat actually is

When the AI labour stack becomes commodity, every founder gets the same starter kit. The differentiators move elsewhere — and this is where the difference between a thought experiment and a real business shows up.

• Vertical know-how

  Domain depth that lets the founder pick problems and judge agent output that generalist competitors cannot. The most defensible moat as the AI tooling itself commoditises.

• Distribution and trust

  An audience, a brand, a relationship with regulators or platforms — anything that gives the founder access to demand the agents would not have on their own. Medvi's GLP-1 demand wave is a distribution story before it is an AI story.

• Proprietary data and context

  Customer data, transaction history, internal evals, and curated examples that cannot be regenerated by a generic agent stack. The thing the agent reads but cannot reproduce.

• Taste

  The ability to judge between equally plausible agent outputs. Cheap to write off; expensive to build; the difference between Medvi and a hundred copycats running the same playbook.

• Compliance edges

  Where rules are unsettled, the founder who reads and operates within them well has an advantage. Where rules are settled, the founder who knows where the lines are has an advantage. Same skill, different game.

Why I am writing this

Frequently asked questions

• What is a zero-human company in 2026?

  A complete business whose execution work — engineering, marketing, support, sales, finance, operations — runs on an organisation of AI agents instead of employees, with one human at the top doing intent, governance, and approvals rather than line-level work. The phrase is also used as a product category for tools like Paperclip that model the company itself as a manageable organisation of agents.

• Is this the same as a one-person company?

  Yes. The two phrases are near-synonyms, used interchangeably in 2026 tech contexts. About 90% of the meaning overlaps; the 10% difference is which side of the same idea the speaker emphasises (the team has gone from employees to software, versus one human runs this). Same canonical examples, same stack, same predictions, same reality check. Both pages on this site cover the same thing.

• What are the canonical zero-human company projects?

  Paperclip is the open-source control plane that models a company as an organisation of agents. Anthropic's Project Vend (2025) and the follow-on Andon Market by Andon Labs (2026) put an LLM in charge of a real business. ClawBank's Manfred has demonstrated an agent self-incorporating a US LLC. Coinbase x402 and AWS Bedrock AgentCore Payments are the payment rails. Coinbase's May 2026 memo introduced 'one-person teams' inside a public company; Shopify's 2025 memo created the cultural prerequisites. Codex App and the Symphony pattern showed how to manage multiple agents through a task tracker. Claude Code is the agentic coding worker most often plugged into the stack. Medvi is the most cited reference case with real numbers and real failure modes.

• Does 'zero-human' mean literally zero humans?

  No, and the most accurate references in 2026 do not claim it. The founder is still a human; so are customers, regulators, and counterparties. The narrower claim is operational: no employees execute the day-to-day work, and humans intervene only at intent, governance, and approvals. The honest description of working examples is 'human-supervised autonomous operations.'

• What does the control plane look like?

  Agent identity and scoped permissions, an org chart of agent roles with reporting lines, monthly budgets with hard stops, heartbeats and scheduled routines, governance with human approval gates, an audit trail of every decision and tool call, payment rails (x402 / AgentCore Payments), and durable memory or institutional knowledge. Paperclip is the canonical open-source implementation of these as one coherent product.

• Has anyone built a one-person billion-dollar company yet?

  Not by mid-2026. Sam Altman has predicted it since 2024; Dario Amodei in 2026 said the first might appear by year-end. Medvi is the most cited reference case with $401M of reported 2025 sales but it is not strictly a one-person operation, and its public failure modes are well documented. The prediction is unvalidated; the trajectory is real.

• What goes wrong in practice?

  Agent hallucinations at customer-facing surfaces (Medvi's first-line support reportedly fabricated drug pricing and products that did not exist). Long-horizon drift where agents lose context across days or weeks. Coordination collapse when ten agents work in parallel without a control plane. Cost blow-ups without enforced budgets. Compliance gaps in industries where the human-in-the-loop requirement is structural rather than technical.

Further reading

• One-Person Company — same shift, organised around the founder workflow
• Zero-human company and one-person company: two names for the same shift
• Reading the agent-organisation stack through Paperclip
• What I see working — and not — in zero-employee operations