Can AI agents run a business?

Yes, if a person keeps the last word. Agents can run a company's work when roles, memory and permissions sit in one repo, and a person reviews every change. That arrangement is company as code, and it is what makes the review possible. An owner's seven-agent experiment left half the agents' own files unupdated and two agents unnamed inside a week.

Do I need to know git to run a company this way?

No. You work in Kortix, and an agent opens a change request for anything it wants to change. The repo is the record, not the interface. What happens between the request and the merge is the agent approval workflow.

How do I correct an agent that has the wrong role?

A role is one agents. block in kortix.yaml, with its prompt file, model and permission. An edit an agent makes to the block opens a change request that a person merges; an edit you make from the dashboard can commit straight to the default branch. The first three roles are week one of the first 30 days.

For teams past the demoRevised 4 November 2026

Run the whole company from one open-source repo

The open-source CompanyOS

A company runs on AI agents once its roles, memory and permissions live in one repo a person approves changes to. Kortix is the open-source CompanyOS: your agents, their skills, your company memory, and every connector in one git repo you own, with the agents working on real cloud computers.

Try Kortix See the loop

Open source · Merge is default-deny for agents · One repo you own

A merged change request in the Kortix repository: pull request 11797, agent questions and permission asks park in Review instead of blocking the session, merged into dev from the session-asks branch, its Files changed view showing the diff to .agents/skills/learnings/MEMORY.md and the new memory entry it records. — Thousands of agents in parallel on one config, each on its own cloud computer.github.com/kortix-ai/CompanyOS/pull/11797

kortix.yaml

# Engineeringeng-triage:  description: Triages every new GitHub issue.  prompt_file: agents/engineering/triage.md  tools: [read, grep, glob, web_search]  connectors: [github, kortix_slack]

Source: https://kortix.com

How the loop hands work back

Three artifacts a person reviews: a change request, a scheduled audit and a drafted reply.

Engineering

A fix arrives as a change request.

The agent reproduces a failing export on its own cloud computer, patches it and opens the change for review.

change-requests/fix-export-timeout.diffOutput

agents/reporting.md

Export runs every hour.

Export runs at 06:00 and retries twice.

+1 -1

Opened by eng-triage, waits for a person

Operations

The Monday audit runs on a schedule.

A cron trigger runs the audit; when it finds a gap it opens a change request instead of fixing it quietly.

schedules/monday-audit.mdOutput

trigger: cron 09:00 Monday
scope: billing and access
flags: 2
next: one change request

Ran once, returned two items

Support

The reply is drafted from company memory.

The agent writes the answer from the account's memory files and leaves it for a person to send.

drafts/northwind-invoice-reply.mdOutput

Field	Value
Account	Northwind
Ask	Copy of the March invoice
Draft	Ready for review

Nothing sent without a person

Two ways to put agents to work

More agents on more machines is still the demo; Kortix runs a company repo you review, OpenClaw an assistant on your own machine.

Criterion

Kortix

OpenClaw

What it is

KortixA company repo you own: agents, skills, memory, connectors.

OpenClawAn AI assistant you message from chat apps.

Where agents run

KortixOn their own cloud computers.

OpenClawOn the machine where you install it.

Who approves a change

KortixA person merges a change request; merge is default-deny for agents.

OpenClawThe owner approves the big calls.

Memory

KortixCompany memory in the git repo you own.

OpenClawMarkdown memory files per agent.

Self-host

KortixSelf-host is free.

OpenClawNo hosted tier; it runs on your machine.

Open source

KortixYes

OpenClawYes

Sources: OpenClaw's own page, and an owner's published seven-agent experiment running OpenClaw, with a CEO, a CTO and five staff on one 4GB VPS. Both read 4 November 2026. Kortix's manifest is quoted from kortix.com.

20,000+ GitHub stars

Open source, and yours to run

What buyers ask before moving the company over

Can AI agents run a business?: Yes, if a person keeps the last word. Agents can run a company's work when roles, memory and permissions sit in one repo, and a person reviews every change. That arrangement is company as code, and it is what makes the review possible. An owner's seven-agent experiment left half the agents' own files unupdated and two agents unnamed inside a week.
Do I need to know git to run a company this way?: No. You work in Kortix, and an agent opens a change request for anything it wants to change. The repo is the record, not the interface. What happens between the request and the merge is the agent approval workflow.
How do I correct an agent that has the wrong role?: A role is one agents.<name> block in kortix.yaml, with its prompt file, model and permission. An edit an agent makes to the block opens a change request that a person merges; an edit you make from the dashboard can commit straight to the default branch. The first three roles are week one of the first 30 days.

Hand the first outcome to an agent this week

One role, one repo, one change request a person reviews. Add the next role when the loop holds.

Try Kortix