Langroid: AI agents that collaborate through messages

An approach inspired by the actor model

Langroid starts from a simple idea: an AI agent is an object you can equip, as needed, with a language model, a vector store and tools. You then wrap it in a Task, and tasks hand work to each other by sending messages, much like the actor model in computer science. You don’t need to know that model to use the framework; the payoff is multi-agent systems whose logic stays readable.

The project dates back to 2023, and its authors used it in a research paper published in 2024 (ML for Healthcare) on pharmacovigilance, built on a multi-agent RAG architecture.

What you can do with it

Langroid ships ready-made specialist agents: one that answers questions from documents (RAG), another that queries a SQL database, and a tool mechanism (ToolMessage) that also produces structured output, handy for pulling key fields out of a contract or a report. An adapter turns the tools of any MCP server into Langroid tools.

Getting started

Install the langroid package with pip, configure a model (an OpenAI key or a local model served by Ollama), then create a chat agent and wrap it in a task. The official Colab quick-start notebook builds a two-agent information-extraction example step by step, with nothing to install on your machine. A separate examples repository covers more advanced cases.

Strengths and limitations

Strengths: a coherent, lightweight architecture, independence from LangChain, support for local models, built-in RAG and SQL agents, and an MIT license.

Limitations: Python only; a smaller community than the big frameworks, so fewer third-party tutorials; and a project run by a small team with invitation-only contributions, which matters if you are choosing for the long term.

Alternatives

CrewAI and AutoGen are well-known Python multi-agent frameworks, LlamaIndex specializes in querying documents, and Atomic Agents also bets on a lightweight design.