How to run an AI agent company that actually works
A chat tab is not a company. Files, two review layers, an inventory, and idle as success — what we keep from agent-company playbooks, on Neuro OS.
Discuss this post in AI
Send a pre-filled prompt to ChatGPT, Claude, Gemini, or Perplexity — get a summary, ask follow-ups, or compare ideas from this guide.
People copy a folder of markdown, name one agent CEO, and call it a company. Then the founder becomes QA: every bug on the live site, every duplicate page, every “done” that was a sketch.
There is a public Paperclip company playbook. It is useful because it was written after things broke, not before a keynote. Pattern Automation does not run Paperclip as the product. We run Neuro OS. The shop floor is different. The failures are the same.
This is our cut of that kind of playbook: what to keep, what to put in git, what never belongs in a chat.
A company is files, not a vibe
Before anyone “starts operating,” the missing file is the future incident.
You need, in some form:
- who we are and who we are not (a soul, a vision — boundaries, not a slogan)
- what already exists (an inventory, or agents will rebuild the landing page)
- how work is allowed to land (commits, branches, a definition of done you can check)
- a heartbeat for the person who delegates, and one for each worker
- tools with names (mail, APIs, paths) so nobody invents a second stack
ChatGPT Projects and memory are folders. They are not this. See how to use ChatGPT professionally.
Two layers before the founder
The principle we keep, in our words: nothing reaches the founder until the worker has checked the work and the manager has checked the worker.
If either layer is a footnote, it will be skipped. A rule at the bottom of a prompt is not a step. Number it. Block on it. Ask for evidence: what did you open, what did you click, what still fails.
The CEO-shaped agent is a quality gate, not a postman. Forwarding another agent’s comment as an assessment is how the founder inherits every defect. If the manager cannot tell a missing page from a CSS nick, they cannot be the gate — list the symptoms they must refuse.
Workers need a self-check that is concrete: the thing exists, it matches the request, there are no placeholders, the format the CMS actually wants is the format you shipped. “Code works” is not done. “Explorations” are not five finals.
Idle is success
Without this sentence in the soul, the CEO invents tasks so nobody looks unemployed. Busywork is how you get a second blog, a third brand name, and a week of tokens.
When the phase backlog is empty, the right move is to say so and wait. An idle department with finished deliverables is the point.
Before creating any task, four questions in any language you like: does this already exist, is it in the phase, can a human check the result, would we still want it if the agent were expensive. If no — do not create it.
Inventory before invention
Every heartbeat that delegates starts by reading what is already there. Product names, owned directories, shipped artefacts. Skip that and you get duplicates, the wrong logo, and a roadmap that fights the last launch.
Corrections from the founder are not a mood. They update the instruction, they get logged, and the next similar task is where you verify the fix. Updating a file and never looking again is how the same mistake ships three times.
Anti-patterns we see anyway
- Manager as postman
- Worker marks done without a self-check
- Same bug, new week, because the loop did not close
- Identity drift: the company quietly becomes “helpful assistant”
- Founder on the live site as the only QA
- Markdown pasted into a system that wanted HTML
- Fifteen options shipped as a design
None of this needs Paperclip specifically. It needs a shop floor you can see. Roles. Review. Heartbeats. Ask before anything that spends money or speaks as the company.
What we run
On Neuro OS the same bones live as org chart, budgets, heartbeats, and files in a repo you own. Copying someone else’s templates is a start. Filling the brackets without a gate is how you get a theatre company.
Read the Paperclip playbook if you want their file names. Then put the company in git and make verification a step, not a hope.
Start at Neuro OS if you want the operating system we sell. The playbook is the lesson. The product is the floor.
