Multi-agent systems · Operating systems · Built in-house

MUVES OS: 34 agents, one review gate, nothing sent on its own.

Built in-house at MUVES. It is the system two founders use to run a company of two, and the working model for what we build for clients.

Custom software, at work.

See the custom software services

Built: A background service and a one-page dashboard that route, draft, review and remember
Stack: Node.jsSQLiteOllama (Qwen, Gemma)ClaudeCodexGmailCalendarGitHub App
Timeline: Built and extended through 2026
Status: In daily use on one office PC
Scale: 34 agents in 9 departments, 62 automated checks

Illustration: one request entering a building of small departments, with a single exit door watched by a person

The problem

Two founders, every job in the company.

Two desks side by side buried under colourful sticky notes, folders and envelopes, each tagged with a different department colour.

Clients, builds, research, follow-ups, money and marketing all land on the same two people. The routine part of that work, first drafts, sorting mail, filing notes, checking numbers, piles up fastest.

We wanted AI to take that routine on without ever letting it speak for us unchecked, and without paying top-model prices for small talk.

What we built

One assistant at the front, a company of agents behind it.

We talk to one assistant, Arthur, by text or voice. Behind him, each request is sorted, routed, drafted, reviewed and remembered.

How it fits together. Cheap decisions happen first and locally. Expensive models only see work that needs them. Anything leaving the company stops twice: at a department head, then at one of us.

A local front door
Small talk is answered on the PC without calling any cloud model. Everything else is sorted by plain rules first and a model second, because rules are free and never guess.
A model router that explains itself
decide.js picks the engine from the live allowance, which engines are up, whether the work is graded, how each model has done lately and which local model is already loaded. It writes down the reason every time. Jobs over 6,000 characters go up a tier automatically.
Departments of agents
34 agents in 9 departments, each with a narrow job and its own instructions. House style is applied before anything is shown.
A review gate
Answers meant to leave the company are held while the department head reviews them. Then they wait for a person. Nothing is ever sent on its own.
Memory that comes from finished work
When a job is done, what the agent learned becomes a note for next time. Agents can also propose small skills, which a reviewer agent accepts, replaces or rejects before any reaches a prompt.
An overseer with no model
A separate process watches the database read-only for stuck, doubled or misfiled work. It has its own case study.

Routing

Route on the cost of being wrong.

Not on the cost of the model. A wrong answer to small talk costs nothing. A wrong email to a client costs a lot.

Tier	Where it runs	Typical use
Local	Qwen and Gemma on the office PC's graphics card	Sorting, tagging, pulling fields out of text. Nothing leaves the building
Codex	Cloud	The next step up, often for code
Smaller Claude	Cloud	Everyday drafting that someone will read
Best Claude	Cloud	Work that is graded, long, or expensive to get wrong

Autonomy

Every workflow starts human-led and has to earn the next step.

Each repeatable job is written down with its steps, the agents that do it, what it replaces, what the person still owns, and where it sits on a three-step ladder. Moving up is a decision we make, not something the system does.

Step 1

Human-led

A person does the work. Agents fetch, sort and prepare.

Step 2

Human-assisted

Agents draft. A person reviews, changes what needs changing, and sends.

Step 3

Autonomous

Agents finish the job and report. Client follow-up is marked "never autonomous" and stays below this line.

How it runs

A normal day with it.

A monitor at sunrise showing a tidy morning brief with calendar, deadlines and one open incident, a coffee steaming beside it.

MorningThe home screen shows a brief built from the calendar, deadlines and any open incidents.
All dayConnected Gmail accounts are checked every 3 minutes using only the sender, subject and preview. Promotions are skipped. Phone alerts say that something needs us, never what it says.
When something stopsThe service restarts itself and the local model engine, with limits: never twice in 10 minutes, at most three times an hour, so a real fault is not hidden by endless restarts.
SpendingUsage is read from the providers themselves, never estimated, and part of each week's Claude allowance is kept back for us.
SecretsAnything shaped like a password or token is stripped before an error is stored or shown, and a scan checks every push for secrets.

What changed

A company of two with a back office.

34agents in 9 departments

12 of 12routing benchmark, about 1 s each

62checks in the main test suite

The honest result is a habit more than a number: one of us asks, the system drafts, one of us approves. We write about running it in How two people run a company with 34 AI agents.

What we'd reuse

A smaller version fits most offices.

Most businesses do not need 34 agents. They need three or four with the same bones: a front door, a router, a review gate and a log.

For	The first agents we would set up
A law or accounting firm	Inbox triage, a document-request drafter, a deadline brief. Client mail always waits for a person
A property manager	Maintenance-request sorting, owner update drafts, a lease-date brief
A young software company	Support triage, release-note drafts, a router that keeps simple questions on a local model