Skip to content

What is MAS-Lab?

The toolkit

MAS Lab builds, runs, and tests multi-agent systems from YAML specifications.

A multi-agent system (MAS) is a team of LLM agents with tools and a workflow that says who talks to whom. MAS Lab is an open toolkit to engineer one.

  1. You describe agents, tools, workflows, and policies in YAML files (manifests).
  2. The runtime runs them and checks behaviour at boundaries such as tool calls and delegations.
  3. Experiments and traces let you compare designs and audit what happened.

The same specification serves local development, benchmarks, and governed deployment.

This page is the toolkit view. For MAS-Lab's place in the Internet of Cognition, start at the landing page.

Why use it

Prototype demos are easy. Production systems are hard to reproduce, debug, and govern. A small change to a prompt, a tool, or the coordination can silently change outcomes.

MAS Lab separates what the system should do (the spec) from what it did (the trace). Contracts, overlays, and observability connect the two.

Who it is for

  • Developers build modular agent apps from specs. Declare agents, tools, and workflows once, then try other models and coordination patterns.
  • Enterprises add governance, traceability, and budgets as overlays. An overlay is a small YAML file merged onto a manifest, so core logic is untouched.
  • Researchers run controlled multi-run experiments. They vary designs and infrastructure, and replay recorded runs.

More on the problem and the design: The problem.

How it is built

The rest of Getting started splits the former architecture page into shorter topics. The five stages, the experiment control plane, the packages and the repository layout stay in this sidebar.

Topic What it covers
The problem Why mixed prompts and glue code fail, who the toolkit is for
Architecture Five stages from declaration to result
Experiments Control plane, design modes, where guarantees come from
Packages Installable Python packages and how they layer
Repo structure Directories in a source checkout
Surfaces mas-ctl, mas-lab, web UI, live dashboard, controller API

Where to start

Browse manifests and runs in a browser with the Web UI. Watch runs live in the live dashboard. For what shipped in this release, read the v0.2 release post.

Reference pages

Topic Where
Install and day-to-day use User guide
Command line CLI index · mas-ctl · mas-lab
Workspace config.yaml config.yaml reference
Labs vs libraries labs-and-libraries.md
Terms Glossary
OpenAI, MCP, A2A, Agent Skills, OpenTelemetry Open standards
YAML manifests, schemas, contracts References
Design-space exploration labs Paper
Releases and updates Blog

Citing this work

If you use MAS Lab in research or publications, cite the MAS Lab article (arXiv:2606.30546), not only this repository. BibTeX and reproduction details are on the Paper page.

Next: Install or Architecture.