DisSysLab — Build an Office of Specialist Agents

AI chatbots answer when you ask. DisSysLab runs an office of specialist agents that works for you continuously — monitoring sources, filtering, analyzing, and delivering results, until you tell it to stop. Each agent has one well-defined job. You describe the office in plain English.

Try it in seconds

pip install dissyslab
dsl run periodic_brief
open brief.html

You'll get a single styled HTML page with three headlines from BBC, three from NPR, current weather for Pasadena, and three stock tickers. No API key, no model download.

dsl list shows every office that ships with DisSysLab. dsl init periodic_brief my_brief copies one into a folder you own so you can edit it.

Mix and match agents to fit your accuracy, budget, and privacy needs

A single office can mix agents that run on different engines: paid AI services (Anthropic Claude, OpenAI, Google Gemini); free local models such as Qwen on Ollama for roles where cost or privacy matters more than accuracy; specialized open-weight models for narrow jobs like coding or entity extraction; and plain Python functions for deterministic tasks. Pick the right engine per agent:

• Ollama — free, local, private. Download a ~19 GB model (Qwen3) one time. Office runs take 15–60 min on a typical Mac.
• OpenRouter — hosted Qwen-2.5-7B. Pennies per run, finishes in 1–5 min on any laptop. API key from openrouter.ai/keys.
• Claude — Anthropic's hosted model. Highest quality, ~25–50¢ per office run. API key from console.anthropic.com.

# Example: OpenRouter, the fast-and-cheap path
export DSL_BACKEND=openrouter
export OPENROUTER_API_KEY=sk-or-v1-...
export OPENROUTER_MODEL=qwen/qwen-2.5-7b-instruct

dsl run situation_room

Or use the one-line installer to be guided through engine choice and shell-rc setup:

curl -sSf https://raw.githubusercontent.com/kmchandy/DisSysLab/main/install.sh | bash

What is an office?

An office is a team of specialist agents with roles. Connections between agents are specified by an network. You specify each role as a job description in English (or as a Python function for deterministic work), and you specify the network in English.

A single role file looks like this:

# Role: analyst

You are a news analyst who reviews articles and forwards items of
political or economic significance to an editor. Exclude celebrity
gossip, sports, and personal opinions.

If the item is relevant, send to editor.
Otherwise send to discard.

That's an agent. Combine a handful of them with sources and sinks in an office.md file, and you have an office that runs continuously.

Offices can contain offices. Each office is a black box — the organization around it only sees what goes in and what comes out. You can build an office whose agents are offices. So you can structure a complex office as a network of sub-offices, just as a university has sub-offices responsible for the medical school, the engineering school, and the college of humanities.

New in v1.6: checkpoint and resume

DisSysLab v1.6 ships an implementation of the Chandy-Lamport distributed snapshot algorithm, adapted for office-shaped systems. With dsl run --snapshot-interval N, the office takes a consistent global checkpoint every N seconds. With dsl run --resume latest, the office restarts from the most recent snapshot and continues without losing or duplicating work. The recovery_demo gallery office demonstrates the protocol on a small Monte Carlo π estimator. See CHANGELOG.md for full release notes.

Learn more

• Full documentation, source, and contributing guide on GitHub.
• The gallery ships ready-to-run offices: periodic_brief (morning HTML brief), situation_room (multi-agent news enrichment), arxiv_radar (arXiv paper rater), kalshi_market_watch (prediction-market briefing), backyard_birds (BirdNET audio classifier), wildlife_watcher (image classifier), and recovery_demo (v1.6 checkpoint-recovery demo). Three React/ FastAPI apps by Caltech undergraduate Nyasha Makaya — a job-hunting assistant, a wardrobe stylist, and a local-events calendar matcher — use dissyslab as a PyPI dependency.
• Visual walk-through in the micro-course.

Requirements

• Python 3.10 or newer.
• macOS or Linux for the shell installer; Windows works for the core framework via pip install.
• An LLM backend if you want to run the agentic offices — see the engine list above. The periodic_brief demo needs none.

License

MIT — see LICENSE on GitHub.