everyone talks about the coming one-person billion-dollar company we’re interested in a different question what if the real breakthrough is not one person running one giant product, but a minimal-human venture factory systematically operating many small products?
NoHum is an early open design project for that question.
It is a repo-native operating system design for minimal-human companies: companies where most recurring work is executed by agents, while humans stay at the level of direction, architecture, boundaries, approvals, and system redesign.
Not one founder, one product. One operating system, many products.
- an early open design project for a minimal-human venture factory
- a public working surface for company ontology, workflows, agent roles, skills, gates, evals, and governance
- a repository of executable organizational modules designed around
Paperclipas the target runtime - a place where contributors can improve prompts, templates, workflows, review rules, and company-building logic through pull requests
- not the live runtime source of truth for the current NoHum company
- not proof that fully autonomous companies already work
- not a pile of prompts pretending to be an operating system
- not a community-managed control plane for live venture, budget, or board decisions
Most AI-company discussion is still framed around one founder using agents to build one giant product.
NoHum explores a different thesis: the interesting breakthrough may be a company that can systematically launch, evaluate, operate, and retire many small software products through a minimal-human operating system.
That means the real design problem is not just better agents. It is better:
- ontology
- workflows
- role boundaries
- tools and skills
- quality gates
- governance
- observability
- economics
This repository should be read as an early design project with uneven module maturity.
The strongest module today is Research.
The studio-wide self-improvement layer is now present as a first-class operating subsystem, but still newer and less battle-tested than Research.
See the full maturity map in MODULES.md.
Research is the most developed reference module in the repo.
It already has:
- a shared
Idea Cardmodel - specialist role decomposition
- source hierarchy and review discipline
- machine-readable decision reasons
- historical memory surfaces
- an explicit venture-selection doctrine, including structured
value delta
Start here:
If you are a founder, operator, or venture-studio person:
If you are a builder or contributor:
If you want the detailed package/runtime layer:
The root COMPANY.md is intentionally a day-1 core and Research import package.
Downstream Product Bet, Build, GTM, and Support roles live in the repository as
optional library surfaces until a separate activation package is created.
The right way to contribute to NoHum is the same way you would contribute to open software:
- improve module contracts
- tighten workflows
- clarify ontology
- improve templates and task surfaces
- strengthen prompts and role boundaries
- add evals, observability, and review rules
- reduce ambiguity and hidden tribal knowledge
The wrong way to contribute:
- generic “super-agent” ideas
- autonomy claims without ontology
- vibes-only prompt tweaks
- changes that blur governance boundaries
- proposals that silently assume access to live runtime data, approvals, or secrets
See CONTRIBUTING.md for the contribution rules.
This repository is for improving how the company should work.
It is not a place where the public directly governs:
- live venture selection
- budget allocation
- board approvals
- production secrets
- customer or payment operations
Those boundaries are documented in GOVERNANCE.md.
- agents/: role specs and manager prompts
- teams/: team structure
- tasks/: runnable workflow units
- skills/: operational skills
- docs/research/: deepest current module
- docs/templates/: canonical artifacts
- docs/observability/: health, eval, and hygiene surfaces
- docs/experiments/: self-improvement experiment memory
- docs/automation/: transition-machine design
- scripts/self-improvement/: deterministic scoring, replay, world-model, and rollback helpers
The current implementation target is Paperclip.
This repo is designed to help shape:
- company structure
- agent roles
- tasks
- knowledge artifacts
- governance surfaces
- workflow execution logic
inside a minimal-human company runtime.
NoHum is best understood today as:
- early
- ambitious
- uneven
- strongest in research
- still opening the downstream company modules in public
That is intentional.
The goal is not to pretend the company is already solved. The goal is to build the operating system in public.
This repository is licensed under the Apache License, Version 2.0.
See LICENSE.