# Luxas

An open-source, multi-agent harness for autonomous scientific research that reads literature, designs and runs experiments, and writes a cited LaTeX report through to a compiled PDF.

Luxas is given a research topic and crawls the literature across sources such as OpenAlex, arXiv and CrossRef, including paywalled venues through an anti-detect browser, then designs and runs experiments with implementation and tests written by sibling agents that stay blind to each other's work. It produces publication-grade figures from the raw results, writes a LaTeX report, submits that report to adversarial content, figure and layout review, and emits a compiled PDF with real citations, running for hours at a time without a human in the loop and recovering from crashes.

The project describes itself as a harness rather than a model: the reasoning comes from models such as Claude and OpenAI's o3, with a one-line redirect to alternatives such as DeepSeek or Kimi. Its own contribution is a durable workspace around that reasoning, including file-backed memory instead of embeddings or a vector store, detached sub-agent processes, an independent-author pattern meant to block agents from rubber-stamping their own work, and deterministic finish gates.

It is built on top of pi-mono's agent-loop and tool-lifecycle primitives, is written in TypeScript for Node, and is released under the license in the repository.

Website: https://luxas.im