This is not a chatbot.

There is no prompt. There is no wait. The office watches your world and acts — while you work, sleep, or travel.

An office of specialist agents. This is what you get — live, on your laptop.

Here's all you wrote.

No code. Two plain English documents.

Plain English. That's it.

Prose for the job. Four lines at the top declaring the outboxes it may send to and the fields it adds — so the office can be checked before it is run.

An office of specialists.
That never clocks out.

You describe the agents. DisSysLab handles everything else.

Connect any agent to any agent.

Each agent does one thing.
Describe it in English.

Job Description.
It's that simple.

English when AI can do the job.
Python when it must be exact.

Most agents are small enough to specify in English. For deterministic work — dedupe by URL, sliding-window RMS, wrap an ML model — use Python. The framework runs both the same way, and both declare their ports: prose in front matter, Python in AgentRoleEntry.

The declarations are why dsl check can find a message addressed to an inbox that does not exist — it reads them without importing anything.

A role is a job description.
Not a person.

Name your agents.
Give each one a role.

Write one line.
Make a connection.

The whole office.
One screen. Plain English.

Pick the AI per agent.

Cheap models for routine work. Premium models for the writing. Free local models when privacy matters. One office can use all three.

Three things.
That's all you write.

Sources and sinks — where data comes from, where results go — you pick from a library. No writing needed.

Loops welcome.

Many frameworks force a one-way pipeline. DisSysLab does not. An agent can feed its upstream, feed itself, or wait for a downstream verdict before continuing.

Reviewer can send work back to Drafter. A DAG can't.

Snapshot. Restart. Continue.

Take a checkpoint while the office is running. After a crash, a reboot, or a code fix, restart from the last snapshot. No work is lost.

Install it.
Two minutes.

--upgrade, not install — on a machine that already has an older one, pip install prints "Requirement already satisfied" and changes nothing. Works on Mac, Linux, and Windows.

See what's already built.

Make a folder that's yours.

This folder is yours. Edit any file. Run it when you want.

Two commands. One second.

dsl draw shows every connection, naming the outbox it leaves by and the inbox it arrives at. dsl check reads the office and reports its structural faults without running it — an inbox nothing writes to, an agent nothing can reach, work that reaches no sink, a message addressed to an inbox that does not exist. dsl checks explains any code it prints.

It is structural, and it stops where structure stops. An office whose diagram is correct can still deadlock. That boundary — what is knowable before you run and what is only knowable after — is the interesting part, and you meet it the first time an office hangs.

Add your API key.

Run it.
Watch it work.

Change a prompt. Re-run.

Edit a role file in plain English. Run the office again. Watch the output change.

You're ready.

Build something that actually works for you — an office that watches for exactly what you care about.